drivers/net/mlx5/mlx5_flow.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2016 6WIND S.A.
   3  * Copyright 2016 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <sys/queue.h>
   7 #include <stdalign.h>
   8 #include <stdint.h>
   9 #include <string.h>
  10
  11 /* Verbs header. */
  12 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  13 #ifdef PEDANTIC
  14 #pragma GCC diagnostic ignored "-Wpedantic"
  15 #endif
  16 #include <infiniband/verbs.h>
  17 #ifdef PEDANTIC
  18 #pragma GCC diagnostic error "-Wpedantic"
  19 #endif
  20
  21 #include <rte_common.h>
  22 #include <rte_ether.h>
  23 #include <rte_eth_ctrl.h>
  24 #include <rte_ethdev_driver.h>
  25 #include <rte_flow.h>
  26 #include <rte_flow_driver.h>
  27 #include <rte_malloc.h>
  28 #include <rte_ip.h>
  29
  30 #include "mlx5.h"
  31 #include "mlx5_defs.h"
  32 #include "mlx5_prm.h"
  33 #include "mlx5_glue.h"
  34
  35 /* Dev ops structure defined in mlx5.c */
  36 extern const struct eth_dev_ops mlx5_dev_ops;
  37 extern const struct eth_dev_ops mlx5_dev_ops_isolate;
  38
  39 /* Pattern outer Layer bits. */
  40 #define MLX5_FLOW_LAYER_OUTER_L2 (1u << 0)
  41 #define MLX5_FLOW_LAYER_OUTER_L3_IPV4 (1u << 1)
  42 #define MLX5_FLOW_LAYER_OUTER_L3_IPV6 (1u << 2)
  43 #define MLX5_FLOW_LAYER_OUTER_L4_UDP (1u << 3)
  44 #define MLX5_FLOW_LAYER_OUTER_L4_TCP (1u << 4)
  45 #define MLX5_FLOW_LAYER_OUTER_VLAN (1u << 5)
  46
  47 /* Pattern inner Layer bits. */
  48 #define MLX5_FLOW_LAYER_INNER_L2 (1u << 6)
  49 #define MLX5_FLOW_LAYER_INNER_L3_IPV4 (1u << 7)
  50 #define MLX5_FLOW_LAYER_INNER_L3_IPV6 (1u << 8)
  51 #define MLX5_FLOW_LAYER_INNER_L4_UDP (1u << 9)
  52 #define MLX5_FLOW_LAYER_INNER_L4_TCP (1u << 10)
  53 #define MLX5_FLOW_LAYER_INNER_VLAN (1u << 11)
  54
  55 /* Pattern tunnel Layer bits. */
  56 #define MLX5_FLOW_LAYER_VXLAN (1u << 12)
  57 #define MLX5_FLOW_LAYER_VXLAN_GPE (1u << 13)
  58 #define MLX5_FLOW_LAYER_GRE (1u << 14)
  59 #define MLX5_FLOW_LAYER_MPLS (1u << 15)
  60
  61 /* Outer Masks. */
  62 #define MLX5_FLOW_LAYER_OUTER_L3 \
  63         (MLX5_FLOW_LAYER_OUTER_L3_IPV4 | MLX5_FLOW_LAYER_OUTER_L3_IPV6)
  64 #define MLX5_FLOW_LAYER_OUTER_L4 \
  65         (MLX5_FLOW_LAYER_OUTER_L4_UDP | MLX5_FLOW_LAYER_OUTER_L4_TCP)
  66 #define MLX5_FLOW_LAYER_OUTER \
  67         (MLX5_FLOW_LAYER_OUTER_L2 | MLX5_FLOW_LAYER_OUTER_L3 | \
  68          MLX5_FLOW_LAYER_OUTER_L4)
  69
  70 /* Tunnel Masks. */
  71 #define MLX5_FLOW_LAYER_TUNNEL \
  72         (MLX5_FLOW_LAYER_VXLAN | MLX5_FLOW_LAYER_VXLAN_GPE | \
  73          MLX5_FLOW_LAYER_GRE | MLX5_FLOW_LAYER_MPLS)
  74
  75 /* Inner Masks. */
  76 #define MLX5_FLOW_LAYER_INNER_L3 \
  77         (MLX5_FLOW_LAYER_INNER_L3_IPV4 | MLX5_FLOW_LAYER_INNER_L3_IPV6)
  78 #define MLX5_FLOW_LAYER_INNER_L4 \
  79         (MLX5_FLOW_LAYER_INNER_L4_UDP | MLX5_FLOW_LAYER_INNER_L4_TCP)
  80 #define MLX5_FLOW_LAYER_INNER \
  81         (MLX5_FLOW_LAYER_INNER_L2 | MLX5_FLOW_LAYER_INNER_L3 | \
  82          MLX5_FLOW_LAYER_INNER_L4)
  83
  84 /* Actions that modify the fate of matching traffic. */
  85 #define MLX5_FLOW_FATE_DROP (1u << 0)
  86 #define MLX5_FLOW_FATE_QUEUE (1u << 1)
  87 #define MLX5_FLOW_FATE_RSS (1u << 2)
  88
  89 /* Modify a packet. */
  90 #define MLX5_FLOW_MOD_FLAG (1u << 0)
  91 #define MLX5_FLOW_MOD_MARK (1u << 1)
  92 #define MLX5_FLOW_MOD_COUNT (1u << 2)
  93
  94 /* possible L3 layers protocols filtering. */
  95 #define MLX5_IP_PROTOCOL_TCP 6
  96 #define MLX5_IP_PROTOCOL_UDP 17
  97 #define MLX5_IP_PROTOCOL_GRE 47
  98 #define MLX5_IP_PROTOCOL_MPLS 147
  99
 100 /* Priority reserved for default flows. */
 101 #define MLX5_FLOW_PRIO_RSVD ((uint32_t)-1)
 102
 103 enum mlx5_expansion {
 104         MLX5_EXPANSION_ROOT,
 105         MLX5_EXPANSION_ROOT_OUTER,
 106         MLX5_EXPANSION_ROOT_ETH_VLAN,
 107         MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN,
 108         MLX5_EXPANSION_OUTER_ETH,
 109         MLX5_EXPANSION_OUTER_ETH_VLAN,
 110         MLX5_EXPANSION_OUTER_VLAN,
 111         MLX5_EXPANSION_OUTER_IPV4,
 112         MLX5_EXPANSION_OUTER_IPV4_UDP,
 113         MLX5_EXPANSION_OUTER_IPV4_TCP,
 114         MLX5_EXPANSION_OUTER_IPV6,
 115         MLX5_EXPANSION_OUTER_IPV6_UDP,
 116         MLX5_EXPANSION_OUTER_IPV6_TCP,
 117         MLX5_EXPANSION_VXLAN,
 118         MLX5_EXPANSION_VXLAN_GPE,
 119         MLX5_EXPANSION_GRE,
 120         MLX5_EXPANSION_MPLS,
 121         MLX5_EXPANSION_ETH,
 122         MLX5_EXPANSION_ETH_VLAN,
 123         MLX5_EXPANSION_VLAN,
 124         MLX5_EXPANSION_IPV4,
 125         MLX5_EXPANSION_IPV4_UDP,
 126         MLX5_EXPANSION_IPV4_TCP,
 127         MLX5_EXPANSION_IPV6,
 128         MLX5_EXPANSION_IPV6_UDP,
 129         MLX5_EXPANSION_IPV6_TCP,
 130 };
 131
 132 /** Supported expansion of items. */
 133 static const struct rte_flow_expand_node mlx5_support_expansion[] = {
 134         [MLX5_EXPANSION_ROOT] = {
 135                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
 136                                                  MLX5_EXPANSION_IPV4,
 137                                                  MLX5_EXPANSION_IPV6),
 138                 .type = RTE_FLOW_ITEM_TYPE_END,
 139         },
 140         [MLX5_EXPANSION_ROOT_OUTER] = {
 141                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_ETH,
 142                                                  MLX5_EXPANSION_OUTER_IPV4,
 143                                                  MLX5_EXPANSION_OUTER_IPV6),
 144                 .type = RTE_FLOW_ITEM_TYPE_END,
 145         },
 146         [MLX5_EXPANSION_ROOT_ETH_VLAN] = {
 147                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH_VLAN),
 148                 .type = RTE_FLOW_ITEM_TYPE_END,
 149         },
 150         [MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN] = {
 151                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_ETH_VLAN),
 152                 .type = RTE_FLOW_ITEM_TYPE_END,
 153         },
 154         [MLX5_EXPANSION_OUTER_ETH] = {
 155                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
 156                                                  MLX5_EXPANSION_OUTER_IPV6,
 157                                                  MLX5_EXPANSION_MPLS),
 158                 .type = RTE_FLOW_ITEM_TYPE_ETH,
 159                 .rss_types = 0,
 160         },
 161         [MLX5_EXPANSION_OUTER_ETH_VLAN] = {
 162                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_VLAN),
 163                 .type = RTE_FLOW_ITEM_TYPE_ETH,
 164                 .rss_types = 0,
 165         },
 166         [MLX5_EXPANSION_OUTER_VLAN] = {
 167                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_OUTER_IPV4,
 168                                                  MLX5_EXPANSION_OUTER_IPV6),
 169                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
 170         },
 171         [MLX5_EXPANSION_OUTER_IPV4] = {
 172                 .next = RTE_FLOW_EXPAND_RSS_NEXT
 173                         (MLX5_EXPANSION_OUTER_IPV4_UDP,
 174                          MLX5_EXPANSION_OUTER_IPV4_TCP,
 175                          MLX5_EXPANSION_GRE),
 176                 .type = RTE_FLOW_ITEM_TYPE_IPV4,
 177                 .rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
 178                         ETH_RSS_NONFRAG_IPV4_OTHER,
 179         },
 180         [MLX5_EXPANSION_OUTER_IPV4_UDP] = {
 181                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
 182                                                  MLX5_EXPANSION_VXLAN_GPE),
 183                 .type = RTE_FLOW_ITEM_TYPE_UDP,
 184                 .rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
 185         },
 186         [MLX5_EXPANSION_OUTER_IPV4_TCP] = {
 187                 .type = RTE_FLOW_ITEM_TYPE_TCP,
 188                 .rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
 189         },
 190         [MLX5_EXPANSION_OUTER_IPV6] = {
 191                 .next = RTE_FLOW_EXPAND_RSS_NEXT
 192                         (MLX5_EXPANSION_OUTER_IPV6_UDP,
 193                          MLX5_EXPANSION_OUTER_IPV6_TCP),
 194                 .type = RTE_FLOW_ITEM_TYPE_IPV6,
 195                 .rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
 196                         ETH_RSS_NONFRAG_IPV6_OTHER,
 197         },
 198         [MLX5_EXPANSION_OUTER_IPV6_UDP] = {
 199                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VXLAN,
 200                                                  MLX5_EXPANSION_VXLAN_GPE),
 201                 .type = RTE_FLOW_ITEM_TYPE_UDP,
 202                 .rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
 203         },
 204         [MLX5_EXPANSION_OUTER_IPV6_TCP] = {
 205                 .type = RTE_FLOW_ITEM_TYPE_TCP,
 206                 .rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
 207         },
 208         [MLX5_EXPANSION_VXLAN] = {
 209                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH),
 210                 .type = RTE_FLOW_ITEM_TYPE_VXLAN,
 211         },
 212         [MLX5_EXPANSION_VXLAN_GPE] = {
 213                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_ETH,
 214                                                  MLX5_EXPANSION_IPV4,
 215                                                  MLX5_EXPANSION_IPV6),
 216                 .type = RTE_FLOW_ITEM_TYPE_VXLAN_GPE,
 217         },
 218         [MLX5_EXPANSION_GRE] = {
 219                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4),
 220                 .type = RTE_FLOW_ITEM_TYPE_GRE,
 221         },
 222         [MLX5_EXPANSION_MPLS] = {
 223                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
 224                                                  MLX5_EXPANSION_IPV6),
 225                 .type = RTE_FLOW_ITEM_TYPE_MPLS,
 226         },
 227         [MLX5_EXPANSION_ETH] = {
 228                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
 229                                                  MLX5_EXPANSION_IPV6),
 230                 .type = RTE_FLOW_ITEM_TYPE_ETH,
 231         },
 232         [MLX5_EXPANSION_ETH_VLAN] = {
 233                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_VLAN),
 234                 .type = RTE_FLOW_ITEM_TYPE_ETH,
 235         },
 236         [MLX5_EXPANSION_VLAN] = {
 237                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4,
 238                                                  MLX5_EXPANSION_IPV6),
 239                 .type = RTE_FLOW_ITEM_TYPE_VLAN,
 240         },
 241         [MLX5_EXPANSION_IPV4] = {
 242                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV4_UDP,
 243                                                  MLX5_EXPANSION_IPV4_TCP),
 244                 .type = RTE_FLOW_ITEM_TYPE_IPV4,
 245                 .rss_types = ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
 246                         ETH_RSS_NONFRAG_IPV4_OTHER,
 247         },
 248         [MLX5_EXPANSION_IPV4_UDP] = {
 249                 .type = RTE_FLOW_ITEM_TYPE_UDP,
 250                 .rss_types = ETH_RSS_NONFRAG_IPV4_UDP,
 251         },
 252         [MLX5_EXPANSION_IPV4_TCP] = {
 253                 .type = RTE_FLOW_ITEM_TYPE_TCP,
 254                 .rss_types = ETH_RSS_NONFRAG_IPV4_TCP,
 255         },
 256         [MLX5_EXPANSION_IPV6] = {
 257                 .next = RTE_FLOW_EXPAND_RSS_NEXT(MLX5_EXPANSION_IPV6_UDP,
 258                                                  MLX5_EXPANSION_IPV6_TCP),
 259                 .type = RTE_FLOW_ITEM_TYPE_IPV6,
 260                 .rss_types = ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
 261                         ETH_RSS_NONFRAG_IPV6_OTHER,
 262         },
 263         [MLX5_EXPANSION_IPV6_UDP] = {
 264                 .type = RTE_FLOW_ITEM_TYPE_UDP,
 265                 .rss_types = ETH_RSS_NONFRAG_IPV6_UDP,
 266         },
 267         [MLX5_EXPANSION_IPV6_TCP] = {
 268                 .type = RTE_FLOW_ITEM_TYPE_TCP,
 269                 .rss_types = ETH_RSS_NONFRAG_IPV6_TCP,
 270         },
 271 };
 272
 273 /** Handles information leading to a drop fate. */
 274 struct mlx5_flow_verbs {
 275         LIST_ENTRY(mlx5_flow_verbs) next;
 276         unsigned int size; /**< Size of the attribute. */
 277         struct {
 278                 struct ibv_flow_attr *attr;
 279                 /**< Pointer to the Specification buffer. */
 280                 uint8_t *specs; /**< Pointer to the specifications. */
 281         };
 282         struct ibv_flow *flow; /**< Verbs flow pointer. */
 283         struct mlx5_hrxq *hrxq; /**< Hash Rx queue object. */
 284         uint64_t hash_fields; /**< Verbs hash Rx queue hash fields. */
 285 };
 286
 287 /* Counters information. */
 288 struct mlx5_flow_counter {
 289         LIST_ENTRY(mlx5_flow_counter) next; /**< Pointer to the next counter. */
 290         uint32_t shared:1; /**< Share counter ID with other flow rules. */
 291         uint32_t ref_cnt:31; /**< Reference counter. */
 292         uint32_t id; /**< Counter ID. */
 293         struct ibv_counter_set *cs; /**< Holds the counters for the rule. */
 294         uint64_t hits; /**< Number of packets matched by the rule. */
 295         uint64_t bytes; /**< Number of bytes matched by the rule. */
 296 };
 297
 298 /* Flow structure. */
 299 struct rte_flow {
 300         TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
 301         struct rte_flow_attr attributes; /**< User flow attribute. */
 302         uint32_t l3_protocol_en:1; /**< Protocol filtering requested. */
 303         uint32_t layers;
 304         /**< Bit-fields of present layers see MLX5_FLOW_LAYER_*. */
 305         uint32_t modifier;
 306         /**< Bit-fields of present modifier see MLX5_FLOW_MOD_*. */
 307         uint32_t fate;
 308         /**< Bit-fields of present fate see MLX5_FLOW_FATE_*. */
 309         uint8_t l3_protocol; /**< valid when l3_protocol_en is set. */
 310         LIST_HEAD(verbs, mlx5_flow_verbs) verbs; /**< Verbs flows list. */
 311         struct mlx5_flow_verbs *cur_verbs;
 312         /**< Current Verbs flow structure being filled. */
 313         struct mlx5_flow_counter *counter; /**< Holds Verbs flow counter. */
 314         struct rte_flow_action_rss rss;/**< RSS context. */
 315         uint8_t key[MLX5_RSS_HASH_KEY_LEN]; /**< RSS hash key. */
 316         uint16_t (*queue)[]; /**< Destination queues to redirect traffic to. */
 317         void *nl_flow; /**< Netlink flow buffer if relevant. */
 318 };
 319
 320 static const struct rte_flow_ops mlx5_flow_ops = {
 321         .validate = mlx5_flow_validate,
 322         .create = mlx5_flow_create,
 323         .destroy = mlx5_flow_destroy,
 324         .flush = mlx5_flow_flush,
 325         .isolate = mlx5_flow_isolate,
 326         .query = mlx5_flow_query,
 327 };
 328
 329 /* Convert FDIR request to Generic flow. */
 330 struct mlx5_fdir {
 331         struct rte_flow_attr attr;
 332         struct rte_flow_action actions[2];
 333         struct rte_flow_item items[4];
 334         struct rte_flow_item_eth l2;
 335         struct rte_flow_item_eth l2_mask;
 336         union {
 337                 struct rte_flow_item_ipv4 ipv4;
 338                 struct rte_flow_item_ipv6 ipv6;
 339         } l3;
 340         union {
 341                 struct rte_flow_item_ipv4 ipv4;
 342                 struct rte_flow_item_ipv6 ipv6;
 343         } l3_mask;
 344         union {
 345                 struct rte_flow_item_udp udp;
 346                 struct rte_flow_item_tcp tcp;
 347         } l4;
 348         union {
 349                 struct rte_flow_item_udp udp;
 350                 struct rte_flow_item_tcp tcp;
 351         } l4_mask;
 352         struct rte_flow_action_queue queue;
 353 };
 354
 355 /* Verbs specification header. */
 356 struct ibv_spec_header {
 357         enum ibv_flow_spec_type type;
 358         uint16_t size;
 359 };
 360
 361 /*
 362  * Number of sub priorities.
 363  * For each kind of pattern matching i.e. L2, L3, L4 to have a correct
 364  * matching on the NIC (firmware dependent) L4 most have the higher priority
 365  * followed by L3 and ending with L2.
 366  */
 367 #define MLX5_PRIORITY_MAP_L2 2
 368 #define MLX5_PRIORITY_MAP_L3 1
 369 #define MLX5_PRIORITY_MAP_L4 0
 370 #define MLX5_PRIORITY_MAP_MAX 3
 371
 372 /* Map of Verbs to Flow priority with 8 Verbs priorities. */
 373 static const uint32_t priority_map_3[][MLX5_PRIORITY_MAP_MAX] = {
 374         { 0, 1, 2 }, { 2, 3, 4 }, { 5, 6, 7 },
 375 };
 376
 377 /* Map of Verbs to Flow priority with 16 Verbs priorities. */
 378 static const uint32_t priority_map_5[][MLX5_PRIORITY_MAP_MAX] = {
 379         { 0, 1, 2 }, { 3, 4, 5 }, { 6, 7, 8 },
 380         { 9, 10, 11 }, { 12, 13, 14 },
 381 };
 382
 383 /* Tunnel information. */
 384 struct mlx5_flow_tunnel_info {
 385         uint32_t tunnel; /**< Tunnel bit (see MLX5_FLOW_*). */
 386         uint32_t ptype; /**< Tunnel Ptype (see RTE_PTYPE_*). */
 387 };
 388
 389 static struct mlx5_flow_tunnel_info tunnels_info[] = {
 390         {
 391                 .tunnel = MLX5_FLOW_LAYER_VXLAN,
 392                 .ptype = RTE_PTYPE_TUNNEL_VXLAN | RTE_PTYPE_L4_UDP,
 393         },
 394         {
 395                 .tunnel = MLX5_FLOW_LAYER_VXLAN_GPE,
 396                 .ptype = RTE_PTYPE_TUNNEL_VXLAN_GPE | RTE_PTYPE_L4_UDP,
 397         },
 398         {
 399                 .tunnel = MLX5_FLOW_LAYER_GRE,
 400                 .ptype = RTE_PTYPE_TUNNEL_GRE,
 401         },
 402         {
 403                 .tunnel = MLX5_FLOW_LAYER_MPLS | MLX5_FLOW_LAYER_OUTER_L4_UDP,
 404                 .ptype = RTE_PTYPE_TUNNEL_MPLS_IN_GRE | RTE_PTYPE_L4_UDP,
 405         },
 406         {
 407                 .tunnel = MLX5_FLOW_LAYER_MPLS,
 408                 .ptype = RTE_PTYPE_TUNNEL_MPLS_IN_GRE,
 409         },
 410 };
 411
 412 /**
 413  * Discover the maximum number of priority available.
 414  *
 415  * @param[in] dev
 416  *   Pointer to Ethernet device.
 417  *
 418  * @return
 419  *   number of supported flow priority on success, a negative errno
 420  *   value otherwise and rte_errno is set.
 421  */
 422 int
 423 mlx5_flow_discover_priorities(struct rte_eth_dev *dev)
 424 {
 425         struct {
 426                 struct ibv_flow_attr attr;
 427                 struct ibv_flow_spec_eth eth;
 428                 struct ibv_flow_spec_action_drop drop;
 429         } flow_attr = {
 430                 .attr = {
 431                         .num_of_specs = 2,
 432                 },
 433                 .eth = {
 434                         .type = IBV_FLOW_SPEC_ETH,
 435                         .size = sizeof(struct ibv_flow_spec_eth),
 436                 },
 437                 .drop = {
 438                         .size = sizeof(struct ibv_flow_spec_action_drop),
 439                         .type = IBV_FLOW_SPEC_ACTION_DROP,
 440                 },
 441         };
 442         struct ibv_flow *flow;
 443         struct mlx5_hrxq *drop = mlx5_hrxq_drop_new(dev);
 444         uint16_t vprio[] = { 8, 16 };
 445         int i;
 446         int priority = 0;
 447
 448         if (!drop) {
 449                 rte_errno = ENOTSUP;
 450                 return -rte_errno;
 451         }
 452         for (i = 0; i != RTE_DIM(vprio); i++) {
 453                 flow_attr.attr.priority = vprio[i] - 1;
 454                 flow = mlx5_glue->create_flow(drop->qp, &flow_attr.attr);
 455                 if (!flow)
 456                         break;
 457                 claim_zero(mlx5_glue->destroy_flow(flow));
 458                 priority = vprio[i];
 459         }
 460         switch (priority) {
 461         case 8:
 462                 priority = RTE_DIM(priority_map_3);
 463                 break;
 464         case 16:
 465                 priority = RTE_DIM(priority_map_5);
 466                 break;
 467         default:
 468                 rte_errno = ENOTSUP;
 469                 DRV_LOG(ERR,
 470                         "port %u verbs maximum priority: %d expected 8/16",
 471                         dev->data->port_id, vprio[i]);
 472                 return -rte_errno;
 473         }
 474         mlx5_hrxq_drop_release(dev);
 475         DRV_LOG(INFO, "port %u flow maximum priority: %d",
 476                 dev->data->port_id, priority);
 477         return priority;
 478 }
 479
 480 /**
 481  * Adjust flow priority.
 482  *
 483  * @param dev
 484  *   Pointer to Ethernet device.
 485  * @param flow
 486  *   Pointer to an rte flow.
 487  */
 488 static void
 489 mlx5_flow_adjust_priority(struct rte_eth_dev *dev, struct rte_flow *flow)
 490 {
 491         struct priv *priv = dev->data->dev_private;
 492         uint32_t priority = flow->attributes.priority;
 493         uint32_t subpriority = flow->cur_verbs->attr->priority;
 494
 495         switch (priv->config.flow_prio) {
 496         case RTE_DIM(priority_map_3):
 497                 priority = priority_map_3[priority][subpriority];
 498                 break;
 499         case RTE_DIM(priority_map_5):
 500                 priority = priority_map_5[priority][subpriority];
 501                 break;
 502         }
 503         flow->cur_verbs->attr->priority = priority;
 504 }
 505
 506 /**
 507  * Get a flow counter.
 508  *
 509  * @param[in] dev
 510  *   Pointer to Ethernet device.
 511  * @param[in] shared
 512  *   Indicate if this counter is shared with other flows.
 513  * @param[in] id
 514  *   Counter identifier.
 515  *
 516  * @return
 517  *   A pointer to the counter, NULL otherwise and rte_errno is set.
 518  */
 519 static struct mlx5_flow_counter *
 520 mlx5_flow_counter_new(struct rte_eth_dev *dev, uint32_t shared, uint32_t id)
 521 {
 522         struct priv *priv = dev->data->dev_private;
 523         struct mlx5_flow_counter *cnt;
 524
 525         LIST_FOREACH(cnt, &priv->flow_counters, next) {
 526                 if (!cnt->shared || cnt->shared != shared)
 527                         continue;
 528                 if (cnt->id != id)
 529                         continue;
 530                 cnt->ref_cnt++;
 531                 return cnt;
 532         }
 533 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 534
 535         struct mlx5_flow_counter tmpl = {
 536                 .shared = shared,
 537                 .id = id,
 538                 .cs = mlx5_glue->create_counter_set
 539                         (priv->ctx,
 540                          &(struct ibv_counter_set_init_attr){
 541                                  .counter_set_id = id,
 542                          }),
 543                 .hits = 0,
 544                 .bytes = 0,
 545         };
 546
 547         if (!tmpl.cs) {
 548                 rte_errno = errno;
 549                 return NULL;
 550         }
 551         cnt = rte_calloc(__func__, 1, sizeof(*cnt), 0);
 552         if (!cnt) {
 553                 rte_errno = ENOMEM;
 554                 return NULL;
 555         }
 556         *cnt = tmpl;
 557         LIST_INSERT_HEAD(&priv->flow_counters, cnt, next);
 558         return cnt;
 559 #endif
 560         rte_errno = ENOTSUP;
 561         return NULL;
 562 }
 563
 564 /**
 565  * Release a flow counter.
 566  *
 567  * @param[in] counter
 568  *   Pointer to the counter handler.
 569  */
 570 static void
 571 mlx5_flow_counter_release(struct mlx5_flow_counter *counter)
 572 {
 573         if (--counter->ref_cnt == 0) {
 574                 claim_zero(mlx5_glue->destroy_counter_set(counter->cs));
 575                 LIST_REMOVE(counter, next);
 576                 rte_free(counter);
 577         }
 578 }
 579
 580 /**
 581  * Verify the @p attributes will be correctly understood by the NIC and store
 582  * them in the @p flow if everything is correct.
 583  *
 584  * @param[in] dev
 585  *   Pointer to Ethernet device.
 586  * @param[in] attributes
 587  *   Pointer to flow attributes
 588  * @param[in, out] flow
 589  *   Pointer to the rte_flow structure.
 590  * @param[out] error
 591  *   Pointer to error structure.
 592  *
 593  * @return
 594  *   0 on success, a negative errno value otherwise and rte_errno is set.
 595  */
 596 static int
 597 mlx5_flow_attributes(struct rte_eth_dev *dev,
 598                      const struct rte_flow_attr *attributes,
 599                      struct rte_flow *flow,
 600                      struct rte_flow_error *error)
 601 {
 602         uint32_t priority_max =
 603                 ((struct priv *)dev->data->dev_private)->config.flow_prio - 1;
 604
 605         if (attributes->group)
 606                 return rte_flow_error_set(error, ENOTSUP,
 607                                           RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 608                                           NULL,
 609                                           "groups is not supported");
 610         if (attributes->priority != MLX5_FLOW_PRIO_RSVD &&
 611             attributes->priority >= priority_max)
 612                 return rte_flow_error_set(error, ENOTSUP,
 613                                           RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 614                                           NULL,
 615                                           "priority out of range");
 616         if (attributes->egress)
 617                 return rte_flow_error_set(error, ENOTSUP,
 618                                           RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 619                                           NULL,
 620                                           "egress is not supported");
 621         if (attributes->transfer)
 622                 return rte_flow_error_set(error, ENOTSUP,
 623                                           RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
 624                                           NULL,
 625                                           "transfer is not supported");
 626         if (!attributes->ingress)
 627                 return rte_flow_error_set(error, ENOTSUP,
 628                                           RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 629                                           NULL,
 630                                           "ingress attribute is mandatory");
 631         flow->attributes = *attributes;
 632         if (attributes->priority == MLX5_FLOW_PRIO_RSVD)
 633                 flow->attributes.priority = priority_max;
 634         return 0;
 635 }
 636
 637 /**
 638  * Verify the @p item specifications (spec, last, mask) are compatible with the
 639  * NIC capabilities.
 640  *
 641  * @param[in] item
 642  *   Item specification.
 643  * @param[in] mask
 644  *   @p item->mask or flow default bit-masks.
 645  * @param[in] nic_mask
 646  *   Bit-masks covering supported fields by the NIC to compare with user mask.
 647  * @param[in] size
 648  *   Bit-masks size in bytes.
 649  * @param[out] error
 650  *   Pointer to error structure.
 651  *
 652  * @return
 653  *   0 on success, a negative errno value otherwise and rte_errno is set.
 654  */
 655 static int
 656 mlx5_flow_item_acceptable(const struct rte_flow_item *item,
 657                           const uint8_t *mask,
 658                           const uint8_t *nic_mask,
 659                           unsigned int size,
 660                           struct rte_flow_error *error)
 661 {
 662         unsigned int i;
 663
 664         assert(nic_mask);
 665         for (i = 0; i < size; ++i)
 666                 if ((nic_mask[i] | mask[i]) != nic_mask[i])
 667                         return rte_flow_error_set(error, ENOTSUP,
 668                                                   RTE_FLOW_ERROR_TYPE_ITEM,
 669                                                   item,
 670                                                   "mask enables non supported"
 671                                                   " bits");
 672         if (!item->spec && (item->mask || item->last))
 673                 return rte_flow_error_set(error, EINVAL,
 674                                           RTE_FLOW_ERROR_TYPE_ITEM,
 675                                           item,
 676                                           "mask/last without a spec is not"
 677                                           " supported");
 678         if (item->spec && item->last) {
 679                 uint8_t spec[size];
 680                 uint8_t last[size];
 681                 unsigned int i;
 682                 int ret;
 683
 684                 for (i = 0; i < size; ++i) {
 685                         spec[i] = ((const uint8_t *)item->spec)[i] & mask[i];
 686                         last[i] = ((const uint8_t *)item->last)[i] & mask[i];
 687                 }
 688                 ret = memcmp(spec, last, size);
 689                 if (ret != 0)
 690                         return rte_flow_error_set(error, ENOTSUP,
 691                                                   RTE_FLOW_ERROR_TYPE_ITEM,
 692                                                   item,
 693                                                   "range is not supported");
 694         }
 695         return 0;
 696 }
 697
 698 /**
 699  * Add a verbs item specification into @p flow.
 700  *
 701  * @param[in, out] flow
 702  *   Pointer to flow structure.
 703  * @param[in] src
 704  *   Create specification.
 705  * @param[in] size
 706  *   Size in bytes of the specification to copy.
 707  */
 708 static void
 709 mlx5_flow_spec_verbs_add(struct rte_flow *flow, void *src, unsigned int size)
 710 {
 711         struct mlx5_flow_verbs *verbs = flow->cur_verbs;
 712
 713         if (verbs->specs) {
 714                 void *dst;
 715
 716                 dst = (void *)(verbs->specs + verbs->size);
 717                 memcpy(dst, src, size);
 718                 ++verbs->attr->num_of_specs;
 719         }
 720         verbs->size += size;
 721 }
 722
 723 /**
 724  * Adjust verbs hash fields according to the @p flow information.
 725  *
 726  * @param[in, out] flow.
 727  *   Pointer to flow structure.
 728  * @param[in] tunnel
 729  *   1 when the hash field is for a tunnel item.
 730  * @param[in] layer_types
 731  *   ETH_RSS_* types.
 732  * @param[in] hash_fields
 733  *   Item hash fields.
 734  */
 735 static void
 736 mlx5_flow_verbs_hashfields_adjust(struct rte_flow *flow,
 737                                   int tunnel __rte_unused,
 738                                   uint32_t layer_types, uint64_t hash_fields)
 739 {
 740 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
 741         hash_fields |= (tunnel ? IBV_RX_HASH_INNER : 0);
 742         if (flow->rss.level == 2 && !tunnel)
 743                 hash_fields = 0;
 744         else if (flow->rss.level < 2 && tunnel)
 745                 hash_fields = 0;
 746 #endif
 747         if (!(flow->rss.types & layer_types))
 748                 hash_fields = 0;
 749         flow->cur_verbs->hash_fields |= hash_fields;
 750 }
 751
 752 /**
 753  * Convert the @p item into a Verbs specification after ensuring the NIC
 754  * will understand and process it correctly.
 755  * If the necessary size for the conversion is greater than the @p flow_size,
 756  * nothing is written in @p flow, the validation is still performed.
 757  *
 758  * @param[in] item
 759  *   Item specification.
 760  * @param[in, out] flow
 761  *   Pointer to flow structure.
 762  * @param[in] flow_size
 763  *   Size in bytes of the available space in @p flow, if too small, nothing is
 764  *   written.
 765  * @param[out] error
 766  *   Pointer to error structure.
 767  *
 768  * @return
 769  *   On success the number of bytes consumed/necessary, if the returned value
 770  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
 771  *   otherwise another call with this returned memory size should be done.
 772  *   On error, a negative errno value is returned and rte_errno is set.
 773  */
 774 static int
 775 mlx5_flow_item_eth(const struct rte_flow_item *item, struct rte_flow *flow,
 776                    const size_t flow_size, struct rte_flow_error *error)
 777 {
 778         const struct rte_flow_item_eth *spec = item->spec;
 779         const struct rte_flow_item_eth *mask = item->mask;
 780         const struct rte_flow_item_eth nic_mask = {
 781                 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 782                 .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 783                 .type = RTE_BE16(0xffff),
 784         };
 785         const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
 786         const unsigned int size = sizeof(struct ibv_flow_spec_eth);
 787         struct ibv_flow_spec_eth eth = {
 788                 .type = IBV_FLOW_SPEC_ETH | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
 789                 .size = size,
 790         };
 791         int ret;
 792
 793         if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
 794                             MLX5_FLOW_LAYER_OUTER_L2))
 795                 return rte_flow_error_set(error, ENOTSUP,
 796                                           RTE_FLOW_ERROR_TYPE_ITEM,
 797                                           item,
 798                                           "L2 layers already configured");
 799         if (!mask)
 800                 mask = &rte_flow_item_eth_mask;
 801         ret = mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
 802                                         (const uint8_t *)&nic_mask,
 803                                         sizeof(struct rte_flow_item_eth),
 804                                         error);
 805         if (ret)
 806                 return ret;
 807         flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
 808                 MLX5_FLOW_LAYER_OUTER_L2;
 809         if (size > flow_size)
 810                 return size;
 811         if (spec) {
 812                 unsigned int i;
 813
 814                 memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 815                 memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 816                 eth.val.ether_type = spec->type;
 817                 memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 818                 memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 819                 eth.mask.ether_type = mask->type;
 820                 /* Remove unwanted bits from values. */
 821                 for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 822                         eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
 823                         eth.val.src_mac[i] &= eth.mask.src_mac[i];
 824                 }
 825                 eth.val.ether_type &= eth.mask.ether_type;
 826         }
 827         flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
 828         mlx5_flow_spec_verbs_add(flow, &eth, size);
 829         return size;
 830 }
 831
 832 /**
 833  * Update the VLAN tag in the Verbs Ethernet specification.
 834  *
 835  * @param[in, out] attr
 836  *   Pointer to Verbs attributes structure.
 837  * @param[in] eth
 838  *   Verbs structure containing the VLAN information to copy.
 839  */
 840 static void
 841 mlx5_flow_item_vlan_update(struct ibv_flow_attr *attr,
 842                            struct ibv_flow_spec_eth *eth)
 843 {
 844         unsigned int i;
 845         const enum ibv_flow_spec_type search = eth->type;
 846         struct ibv_spec_header *hdr = (struct ibv_spec_header *)
 847                 ((uint8_t *)attr + sizeof(struct ibv_flow_attr));
 848
 849         for (i = 0; i != attr->num_of_specs; ++i) {
 850                 if (hdr->type == search) {
 851                         struct ibv_flow_spec_eth *e =
 852                                 (struct ibv_flow_spec_eth *)hdr;
 853
 854                         e->val.vlan_tag = eth->val.vlan_tag;
 855                         e->mask.vlan_tag = eth->mask.vlan_tag;
 856                         e->val.ether_type = eth->val.ether_type;
 857                         e->mask.ether_type = eth->mask.ether_type;
 858                         break;
 859                 }
 860                 hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
 861         }
 862 }
 863
 864 /**
 865  * Convert the @p item into @p flow (or by updating the already present
 866  * Ethernet Verbs) specification after ensuring the NIC will understand and
 867  * process it correctly.
 868  * If the necessary size for the conversion is greater than the @p flow_size,
 869  * nothing is written in @p flow, the validation is still performed.
 870  *
 871  * @param[in] item
 872  *   Item specification.
 873  * @param[in, out] flow
 874  *   Pointer to flow structure.
 875  * @param[in] flow_size
 876  *   Size in bytes of the available space in @p flow, if too small, nothing is
 877  *   written.
 878  * @param[out] error
 879  *   Pointer to error structure.
 880  *
 881  * @return
 882  *   On success the number of bytes consumed/necessary, if the returned value
 883  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
 884  *   otherwise another call with this returned memory size should be done.
 885  *   On error, a negative errno value is returned and rte_errno is set.
 886  */
 887 static int
 888 mlx5_flow_item_vlan(const struct rte_flow_item *item, struct rte_flow *flow,
 889                     const size_t flow_size, struct rte_flow_error *error)
 890 {
 891         const struct rte_flow_item_vlan *spec = item->spec;
 892         const struct rte_flow_item_vlan *mask = item->mask;
 893         const struct rte_flow_item_vlan nic_mask = {
 894                 .tci = RTE_BE16(0x0fff),
 895                 .inner_type = RTE_BE16(0xffff),
 896         };
 897         unsigned int size = sizeof(struct ibv_flow_spec_eth);
 898         const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
 899         struct ibv_flow_spec_eth eth = {
 900                 .type = IBV_FLOW_SPEC_ETH | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
 901                 .size = size,
 902         };
 903         int ret;
 904         const uint32_t l34m = tunnel ? (MLX5_FLOW_LAYER_INNER_L3 |
 905                                         MLX5_FLOW_LAYER_INNER_L4) :
 906                 (MLX5_FLOW_LAYER_OUTER_L3 | MLX5_FLOW_LAYER_OUTER_L4);
 907         const uint32_t vlanm = tunnel ? MLX5_FLOW_LAYER_INNER_VLAN :
 908                 MLX5_FLOW_LAYER_OUTER_VLAN;
 909         const uint32_t l2m = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
 910                 MLX5_FLOW_LAYER_OUTER_L2;
 911
 912         if (flow->layers & vlanm)
 913                 return rte_flow_error_set(error, ENOTSUP,
 914                                           RTE_FLOW_ERROR_TYPE_ITEM,
 915                                           item,
 916                                           "VLAN layer already configured");
 917         else if ((flow->layers & l34m) != 0)
 918                 return rte_flow_error_set(error, ENOTSUP,
 919                                           RTE_FLOW_ERROR_TYPE_ITEM,
 920                                           item,
 921                                           "L2 layer cannot follow L3/L4 layer");
 922         if (!mask)
 923                 mask = &rte_flow_item_vlan_mask;
 924         ret = mlx5_flow_item_acceptable
 925                 (item, (const uint8_t *)mask,
 926                  (const uint8_t *)&nic_mask,
 927                  sizeof(struct rte_flow_item_vlan), error);
 928         if (ret)
 929                 return ret;
 930         if (spec) {
 931                 eth.val.vlan_tag = spec->tci;
 932                 eth.mask.vlan_tag = mask->tci;
 933                 eth.val.vlan_tag &= eth.mask.vlan_tag;
 934                 eth.val.ether_type = spec->inner_type;
 935                 eth.mask.ether_type = mask->inner_type;
 936                 eth.val.ether_type &= eth.mask.ether_type;
 937         }
 938         /*
 939          * From verbs perspective an empty VLAN is equivalent
 940          * to a packet without VLAN layer.
 941          */
 942         if (!eth.mask.vlan_tag)
 943                 return rte_flow_error_set(error, EINVAL,
 944                                           RTE_FLOW_ERROR_TYPE_ITEM_SPEC,
 945                                           item->spec,
 946                                           "VLAN cannot be empty");
 947         if (!(flow->layers & l2m)) {
 948                 if (size <= flow_size) {
 949                         flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
 950                         mlx5_flow_spec_verbs_add(flow, &eth, size);
 951                 }
 952         } else {
 953                 if (flow->cur_verbs)
 954                         mlx5_flow_item_vlan_update(flow->cur_verbs->attr,
 955                                                    &eth);
 956                 size = 0; /* Only an update is done in eth specification. */
 957         }
 958         flow->layers |= tunnel ?
 959                 (MLX5_FLOW_LAYER_INNER_L2 | MLX5_FLOW_LAYER_INNER_VLAN) :
 960                 (MLX5_FLOW_LAYER_OUTER_L2 | MLX5_FLOW_LAYER_OUTER_VLAN);
 961         return size;
 962 }
 963
 964 /**
 965  * Convert the @p item into a Verbs specification after ensuring the NIC
 966  * will understand and process it correctly.
 967  * If the necessary size for the conversion is greater than the @p flow_size,
 968  * nothing is written in @p flow, the validation is still performed.
 969  *
 970  * @param[in] item
 971  *   Item specification.
 972  * @param[in, out] flow
 973  *   Pointer to flow structure.
 974  * @param[in] flow_size
 975  *   Size in bytes of the available space in @p flow, if too small, nothing is
 976  *   written.
 977  * @param[out] error
 978  *   Pointer to error structure.
 979  *
 980  * @return
 981  *   On success the number of bytes consumed/necessary, if the returned value
 982  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
 983  *   otherwise another call with this returned memory size should be done.
 984  *   On error, a negative errno value is returned and rte_errno is set.
 985  */
 986 static int
 987 mlx5_flow_item_ipv4(const struct rte_flow_item *item, struct rte_flow *flow,
 988                     const size_t flow_size, struct rte_flow_error *error)
 989 {
 990         const struct rte_flow_item_ipv4 *spec = item->spec;
 991         const struct rte_flow_item_ipv4 *mask = item->mask;
 992         const struct rte_flow_item_ipv4 nic_mask = {
 993                 .hdr = {
 994                         .src_addr = RTE_BE32(0xffffffff),
 995                         .dst_addr = RTE_BE32(0xffffffff),
 996                         .type_of_service = 0xff,
 997                         .next_proto_id = 0xff,
 998                 },
 999         };
1000         const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1001         unsigned int size = sizeof(struct ibv_flow_spec_ipv4_ext);
1002         struct ibv_flow_spec_ipv4_ext ipv4 = {
1003                 .type = IBV_FLOW_SPEC_IPV4_EXT |
1004                         (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1005                 .size = size,
1006         };
1007         int ret;
1008
1009         if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1010                             MLX5_FLOW_LAYER_OUTER_L3))
1011                 return rte_flow_error_set(error, ENOTSUP,
1012                                           RTE_FLOW_ERROR_TYPE_ITEM,
1013                                           item,
1014                                           "multiple L3 layers not supported");
1015         else if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1016                                  MLX5_FLOW_LAYER_OUTER_L4))
1017                 return rte_flow_error_set(error, ENOTSUP,
1018                                           RTE_FLOW_ERROR_TYPE_ITEM,
1019                                           item,
1020                                           "L3 cannot follow an L4 layer.");
1021         if (!mask)
1022                 mask = &rte_flow_item_ipv4_mask;
1023         ret = mlx5_flow_item_acceptable
1024                 (item, (const uint8_t *)mask,
1025                  (const uint8_t *)&nic_mask,
1026                  sizeof(struct rte_flow_item_ipv4), error);
1027         if (ret < 0)
1028                 return ret;
1029         flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1030                 MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1031         if (spec) {
1032                 ipv4.val = (struct ibv_flow_ipv4_ext_filter){
1033                         .src_ip = spec->hdr.src_addr,
1034                         .dst_ip = spec->hdr.dst_addr,
1035                         .proto = spec->hdr.next_proto_id,
1036                         .tos = spec->hdr.type_of_service,
1037                 };
1038                 ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1039                         .src_ip = mask->hdr.src_addr,
1040                         .dst_ip = mask->hdr.dst_addr,
1041                         .proto = mask->hdr.next_proto_id,
1042                         .tos = mask->hdr.type_of_service,
1043                 };
1044                 /* Remove unwanted bits from values. */
1045                 ipv4.val.src_ip &= ipv4.mask.src_ip;
1046                 ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1047                 ipv4.val.proto &= ipv4.mask.proto;
1048                 ipv4.val.tos &= ipv4.mask.tos;
1049         }
1050         flow->l3_protocol_en = !!ipv4.mask.proto;
1051         flow->l3_protocol = ipv4.val.proto;
1052         if (size <= flow_size) {
1053                 mlx5_flow_verbs_hashfields_adjust
1054                         (flow, tunnel,
1055                          (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
1056                           ETH_RSS_NONFRAG_IPV4_OTHER),
1057                          (IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4));
1058                 flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L3;
1059                 mlx5_flow_spec_verbs_add(flow, &ipv4, size);
1060         }
1061         return size;
1062 }
1063
1064 /**
1065  * Convert the @p item into a Verbs specification after ensuring the NIC
1066  * will understand and process it correctly.
1067  * If the necessary size for the conversion is greater than the @p flow_size,
1068  * nothing is written in @p flow, the validation is still performed.
1069  *
1070  * @param[in] item
1071  *   Item specification.
1072  * @param[in, out] flow
1073  *   Pointer to flow structure.
1074  * @param[in] flow_size
1075  *   Size in bytes of the available space in @p flow, if too small, nothing is
1076  *   written.
1077  * @param[out] error
1078  *   Pointer to error structure.
1079  *
1080  * @return
1081  *   On success the number of bytes consumed/necessary, if the returned value
1082  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1083  *   otherwise another call with this returned memory size should be done.
1084  *   On error, a negative errno value is returned and rte_errno is set.
1085  */
1086 static int
1087 mlx5_flow_item_ipv6(const struct rte_flow_item *item, struct rte_flow *flow,
1088                     const size_t flow_size, struct rte_flow_error *error)
1089 {
1090         const struct rte_flow_item_ipv6 *spec = item->spec;
1091         const struct rte_flow_item_ipv6 *mask = item->mask;
1092         const struct rte_flow_item_ipv6 nic_mask = {
1093                 .hdr = {
1094                         .src_addr =
1095                                 "\xff\xff\xff\xff\xff\xff\xff\xff"
1096                                 "\xff\xff\xff\xff\xff\xff\xff\xff",
1097                         .dst_addr =
1098                                 "\xff\xff\xff\xff\xff\xff\xff\xff"
1099                                 "\xff\xff\xff\xff\xff\xff\xff\xff",
1100                         .vtc_flow = RTE_BE32(0xffffffff),
1101                         .proto = 0xff,
1102                         .hop_limits = 0xff,
1103                 },
1104         };
1105         const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1106         unsigned int size = sizeof(struct ibv_flow_spec_ipv6);
1107         struct ibv_flow_spec_ipv6 ipv6 = {
1108                 .type = IBV_FLOW_SPEC_IPV6 | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1109                 .size = size,
1110         };
1111         int ret;
1112
1113         if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1114                             MLX5_FLOW_LAYER_OUTER_L3))
1115                 return rte_flow_error_set(error, ENOTSUP,
1116                                           RTE_FLOW_ERROR_TYPE_ITEM,
1117                                           item,
1118                                           "multiple L3 layers not supported");
1119         else if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1120                                  MLX5_FLOW_LAYER_OUTER_L4))
1121                 return rte_flow_error_set(error, ENOTSUP,
1122                                           RTE_FLOW_ERROR_TYPE_ITEM,
1123                                           item,
1124                                           "L3 cannot follow an L4 layer.");
1125         /*
1126          * IPv6 is not recognised by the NIC inside a GRE tunnel.
1127          * Such support has to be disabled as the rule will be
1128          * accepted.  Issue reproduced with Mellanox OFED 4.3-3.0.2.1 and
1129          * Mellanox OFED 4.4-1.0.0.0.
1130          */
1131         if (tunnel && flow->layers & MLX5_FLOW_LAYER_GRE)
1132                 return rte_flow_error_set(error, ENOTSUP,
1133                                           RTE_FLOW_ERROR_TYPE_ITEM,
1134                                           item,
1135                                           "IPv6 inside a GRE tunnel is"
1136                                           " not recognised.");
1137         if (!mask)
1138                 mask = &rte_flow_item_ipv6_mask;
1139         ret = mlx5_flow_item_acceptable
1140                 (item, (const uint8_t *)mask,
1141                  (const uint8_t *)&nic_mask,
1142                  sizeof(struct rte_flow_item_ipv6), error);
1143         if (ret < 0)
1144                 return ret;
1145         flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1146                 MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1147         if (spec) {
1148                 unsigned int i;
1149                 uint32_t vtc_flow_val;
1150                 uint32_t vtc_flow_mask;
1151
1152                 memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1153                        RTE_DIM(ipv6.val.src_ip));
1154                 memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1155                        RTE_DIM(ipv6.val.dst_ip));
1156                 memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1157                        RTE_DIM(ipv6.mask.src_ip));
1158                 memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1159                        RTE_DIM(ipv6.mask.dst_ip));
1160                 vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
1161                 vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
1162                 ipv6.val.flow_label =
1163                         rte_cpu_to_be_32((vtc_flow_val & IPV6_HDR_FL_MASK) >>
1164                                          IPV6_HDR_FL_SHIFT);
1165                 ipv6.val.traffic_class = (vtc_flow_val & IPV6_HDR_TC_MASK) >>
1166                                          IPV6_HDR_TC_SHIFT;
1167                 ipv6.val.next_hdr = spec->hdr.proto;
1168                 ipv6.val.hop_limit = spec->hdr.hop_limits;
1169                 ipv6.mask.flow_label =
1170                         rte_cpu_to_be_32((vtc_flow_mask & IPV6_HDR_FL_MASK) >>
1171                                          IPV6_HDR_FL_SHIFT);
1172                 ipv6.mask.traffic_class = (vtc_flow_mask & IPV6_HDR_TC_MASK) >>
1173                                           IPV6_HDR_TC_SHIFT;
1174                 ipv6.mask.next_hdr = mask->hdr.proto;
1175                 ipv6.mask.hop_limit = mask->hdr.hop_limits;
1176                 /* Remove unwanted bits from values. */
1177                 for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1178                         ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1179                         ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1180                 }
1181                 ipv6.val.flow_label &= ipv6.mask.flow_label;
1182                 ipv6.val.traffic_class &= ipv6.mask.traffic_class;
1183                 ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1184                 ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1185         }
1186         flow->l3_protocol_en = !!ipv6.mask.next_hdr;
1187         flow->l3_protocol = ipv6.val.next_hdr;
1188         if (size <= flow_size) {
1189                 mlx5_flow_verbs_hashfields_adjust
1190                         (flow, tunnel,
1191                          (ETH_RSS_IPV6 | ETH_RSS_NONFRAG_IPV6_OTHER),
1192                          (IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6));
1193                 flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L3;
1194                 mlx5_flow_spec_verbs_add(flow, &ipv6, size);
1195         }
1196         return size;
1197 }
1198
1199 /**
1200  * Convert the @p item into a Verbs specification after ensuring the NIC
1201  * will understand and process it correctly.
1202  * If the necessary size for the conversion is greater than the @p flow_size,
1203  * nothing is written in @p flow, the validation is still performed.
1204  *
1205  * @param[in] item
1206  *   Item specification.
1207  * @param[in, out] flow
1208  *   Pointer to flow structure.
1209  * @param[in] flow_size
1210  *   Size in bytes of the available space in @p flow, if too small, nothing is
1211  *   written.
1212  * @param[out] error
1213  *   Pointer to error structure.
1214  *
1215  * @return
1216  *   On success the number of bytes consumed/necessary, if the returned value
1217  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1218  *   otherwise another call with this returned memory size should be done.
1219  *   On error, a negative errno value is returned and rte_errno is set.
1220  */
1221 static int
1222 mlx5_flow_item_udp(const struct rte_flow_item *item, struct rte_flow *flow,
1223                    const size_t flow_size, struct rte_flow_error *error)
1224 {
1225         const struct rte_flow_item_udp *spec = item->spec;
1226         const struct rte_flow_item_udp *mask = item->mask;
1227         const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1228         unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
1229         struct ibv_flow_spec_tcp_udp udp = {
1230                 .type = IBV_FLOW_SPEC_UDP | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1231                 .size = size,
1232         };
1233         int ret;
1234
1235         if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_UDP)
1236                 return rte_flow_error_set(error, ENOTSUP,
1237                                           RTE_FLOW_ERROR_TYPE_ITEM,
1238                                           item,
1239                                           "protocol filtering not compatible"
1240                                           " with UDP layer");
1241         if (!(flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1242                               MLX5_FLOW_LAYER_OUTER_L3)))
1243                 return rte_flow_error_set(error, ENOTSUP,
1244                                           RTE_FLOW_ERROR_TYPE_ITEM,
1245                                           item,
1246                                           "L3 is mandatory to filter"
1247                                           " on L4");
1248         if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1249                             MLX5_FLOW_LAYER_OUTER_L4))
1250                 return rte_flow_error_set(error, ENOTSUP,
1251                                           RTE_FLOW_ERROR_TYPE_ITEM,
1252                                           item,
1253                                           "L4 layer is already"
1254                                           " present");
1255         if (!mask)
1256                 mask = &rte_flow_item_udp_mask;
1257         ret = mlx5_flow_item_acceptable
1258                 (item, (const uint8_t *)mask,
1259                  (const uint8_t *)&rte_flow_item_udp_mask,
1260                  sizeof(struct rte_flow_item_udp), error);
1261         if (ret < 0)
1262                 return ret;
1263         flow->layers |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1264                 MLX5_FLOW_LAYER_OUTER_L4_UDP;
1265         if (spec) {
1266                 udp.val.dst_port = spec->hdr.dst_port;
1267                 udp.val.src_port = spec->hdr.src_port;
1268                 udp.mask.dst_port = mask->hdr.dst_port;
1269                 udp.mask.src_port = mask->hdr.src_port;
1270                 /* Remove unwanted bits from values. */
1271                 udp.val.src_port &= udp.mask.src_port;
1272                 udp.val.dst_port &= udp.mask.dst_port;
1273         }
1274         if (size <= flow_size) {
1275                 mlx5_flow_verbs_hashfields_adjust(flow, tunnel, ETH_RSS_UDP,
1276                                                   (IBV_RX_HASH_SRC_PORT_UDP |
1277                                                    IBV_RX_HASH_DST_PORT_UDP));
1278                 flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L4;
1279                 mlx5_flow_spec_verbs_add(flow, &udp, size);
1280         }
1281         return size;
1282 }
1283
1284 /**
1285  * Convert the @p item into a Verbs specification after ensuring the NIC
1286  * will understand and process it correctly.
1287  * If the necessary size for the conversion is greater than the @p flow_size,
1288  * nothing is written in @p flow, the validation is still performed.
1289  *
1290  * @param[in] item
1291  *   Item specification.
1292  * @param[in, out] flow
1293  *   Pointer to flow structure.
1294  * @param[in] flow_size
1295  *   Size in bytes of the available space in @p flow, if too small, nothing is
1296  *   written.
1297  * @param[out] error
1298  *   Pointer to error structure.
1299  *
1300  * @return
1301  *   On success the number of bytes consumed/necessary, if the returned value
1302  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1303  *   otherwise another call with this returned memory size should be done.
1304  *   On error, a negative errno value is returned and rte_errno is set.
1305  */
1306 static int
1307 mlx5_flow_item_tcp(const struct rte_flow_item *item, struct rte_flow *flow,
1308                    const size_t flow_size, struct rte_flow_error *error)
1309 {
1310         const struct rte_flow_item_tcp *spec = item->spec;
1311         const struct rte_flow_item_tcp *mask = item->mask;
1312         const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
1313         unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
1314         struct ibv_flow_spec_tcp_udp tcp = {
1315                 .type = IBV_FLOW_SPEC_TCP | (tunnel ? IBV_FLOW_SPEC_INNER : 0),
1316                 .size = size,
1317         };
1318         int ret;
1319
1320         if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_TCP)
1321                 return rte_flow_error_set(error, ENOTSUP,
1322                                           RTE_FLOW_ERROR_TYPE_ITEM,
1323                                           item,
1324                                           "protocol filtering not compatible"
1325                                           " with TCP layer");
1326         if (!(flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L3 :
1327                               MLX5_FLOW_LAYER_OUTER_L3)))
1328                 return rte_flow_error_set(error, ENOTSUP,
1329                                           RTE_FLOW_ERROR_TYPE_ITEM,
1330                                           item,
1331                                           "L3 is mandatory to filter on L4");
1332         if (flow->layers & (tunnel ? MLX5_FLOW_LAYER_INNER_L4 :
1333                             MLX5_FLOW_LAYER_OUTER_L4))
1334                 return rte_flow_error_set(error, ENOTSUP,
1335                                           RTE_FLOW_ERROR_TYPE_ITEM,
1336                                           item,
1337                                           "L4 layer is already present");
1338         if (!mask)
1339                 mask = &rte_flow_item_tcp_mask;
1340         ret = mlx5_flow_item_acceptable
1341                 (item, (const uint8_t *)mask,
1342                  (const uint8_t *)&rte_flow_item_tcp_mask,
1343                  sizeof(struct rte_flow_item_tcp), error);
1344         if (ret < 0)
1345                 return ret;
1346         flow->layers |=  tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1347                 MLX5_FLOW_LAYER_OUTER_L4_TCP;
1348         if (spec) {
1349                 tcp.val.dst_port = spec->hdr.dst_port;
1350                 tcp.val.src_port = spec->hdr.src_port;
1351                 tcp.mask.dst_port = mask->hdr.dst_port;
1352                 tcp.mask.src_port = mask->hdr.src_port;
1353                 /* Remove unwanted bits from values. */
1354                 tcp.val.src_port &= tcp.mask.src_port;
1355                 tcp.val.dst_port &= tcp.mask.dst_port;
1356         }
1357         if (size <= flow_size) {
1358                 mlx5_flow_verbs_hashfields_adjust(flow, tunnel, ETH_RSS_TCP,
1359                                                   (IBV_RX_HASH_SRC_PORT_TCP |
1360                                                    IBV_RX_HASH_DST_PORT_TCP));
1361                 flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L4;
1362                 mlx5_flow_spec_verbs_add(flow, &tcp, size);
1363         }
1364         return size;
1365 }
1366
1367 /**
1368  * Convert the @p item into a Verbs specification after ensuring the NIC
1369  * will understand and process it correctly.
1370  * If the necessary size for the conversion is greater than the @p flow_size,
1371  * nothing is written in @p flow, the validation is still performed.
1372  *
1373  * @param[in] item
1374  *   Item specification.
1375  * @param[in, out] flow
1376  *   Pointer to flow structure.
1377  * @param[in] flow_size
1378  *   Size in bytes of the available space in @p flow, if too small, nothing is
1379  *   written.
1380  * @param[out] error
1381  *   Pointer to error structure.
1382  *
1383  * @return
1384  *   On success the number of bytes consumed/necessary, if the returned value
1385  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1386  *   otherwise another call with this returned memory size should be done.
1387  *   On error, a negative errno value is returned and rte_errno is set.
1388  */
1389 static int
1390 mlx5_flow_item_vxlan(const struct rte_flow_item *item, struct rte_flow *flow,
1391                      const size_t flow_size, struct rte_flow_error *error)
1392 {
1393         const struct rte_flow_item_vxlan *spec = item->spec;
1394         const struct rte_flow_item_vxlan *mask = item->mask;
1395         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1396         struct ibv_flow_spec_tunnel vxlan = {
1397                 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1398                 .size = size,
1399         };
1400         int ret;
1401         union vni {
1402                 uint32_t vlan_id;
1403                 uint8_t vni[4];
1404         } id = { .vlan_id = 0, };
1405
1406         if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1407                 return rte_flow_error_set(error, ENOTSUP,
1408                                           RTE_FLOW_ERROR_TYPE_ITEM,
1409                                           item,
1410                                           "a tunnel is already present");
1411         /*
1412          * Verify only UDPv4 is present as defined in
1413          * https://tools.ietf.org/html/rfc7348
1414          */
1415         if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L4_UDP))
1416                 return rte_flow_error_set(error, ENOTSUP,
1417                                           RTE_FLOW_ERROR_TYPE_ITEM,
1418                                           item,
1419                                           "no outer UDP layer found");
1420         if (!mask)
1421                 mask = &rte_flow_item_vxlan_mask;
1422         ret = mlx5_flow_item_acceptable
1423                 (item, (const uint8_t *)mask,
1424                  (const uint8_t *)&rte_flow_item_vxlan_mask,
1425                  sizeof(struct rte_flow_item_vxlan), error);
1426         if (ret < 0)
1427                 return ret;
1428         if (spec) {
1429                 memcpy(&id.vni[1], spec->vni, 3);
1430                 vxlan.val.tunnel_id = id.vlan_id;
1431                 memcpy(&id.vni[1], mask->vni, 3);
1432                 vxlan.mask.tunnel_id = id.vlan_id;
1433                 /* Remove unwanted bits from values. */
1434                 vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1435         }
1436         /*
1437          * Tunnel id 0 is equivalent as not adding a VXLAN layer, if
1438          * only this layer is defined in the Verbs specification it is
1439          * interpreted as wildcard and all packets will match this
1440          * rule, if it follows a full stack layer (ex: eth / ipv4 /
1441          * udp), all packets matching the layers before will also
1442          * match this rule.  To avoid such situation, VNI 0 is
1443          * currently refused.
1444          */
1445         if (!vxlan.val.tunnel_id)
1446                 return rte_flow_error_set(error, EINVAL,
1447                                           RTE_FLOW_ERROR_TYPE_ITEM,
1448                                           item,
1449                                           "VXLAN vni cannot be 0");
1450         if (!(flow->layers & MLX5_FLOW_LAYER_OUTER))
1451                 return rte_flow_error_set(error, EINVAL,
1452                                           RTE_FLOW_ERROR_TYPE_ITEM,
1453                                           item,
1454                                           "VXLAN tunnel must be fully defined");
1455         if (size <= flow_size) {
1456                 mlx5_flow_spec_verbs_add(flow, &vxlan, size);
1457                 flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1458         }
1459         flow->layers |= MLX5_FLOW_LAYER_VXLAN;
1460         return size;
1461 }
1462
1463 /**
1464  * Convert the @p item into a Verbs specification after ensuring the NIC
1465  * will understand and process it correctly.
1466  * If the necessary size for the conversion is greater than the @p flow_size,
1467  * nothing is written in @p flow, the validation is still performed.
1468  *
1469  * @param dev
1470  *   Pointer to Ethernet device.
1471  * @param[in] item
1472  *   Item specification.
1473  * @param[in, out] flow
1474  *   Pointer to flow structure.
1475  * @param[in] flow_size
1476  *   Size in bytes of the available space in @p flow, if too small, nothing is
1477  *   written.
1478  * @param[out] error
1479  *   Pointer to error structure.
1480  *
1481  * @return
1482  *   On success the number of bytes consumed/necessary, if the returned value
1483  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1484  *   otherwise another call with this returned memory size should be done.
1485  *   On error, a negative errno value is returned and rte_errno is set.
1486  */
1487 static int
1488 mlx5_flow_item_vxlan_gpe(struct rte_eth_dev *dev,
1489                          const struct rte_flow_item *item,
1490                          struct rte_flow *flow, const size_t flow_size,
1491                          struct rte_flow_error *error)
1492 {
1493         const struct rte_flow_item_vxlan_gpe *spec = item->spec;
1494         const struct rte_flow_item_vxlan_gpe *mask = item->mask;
1495         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1496         struct ibv_flow_spec_tunnel vxlan_gpe = {
1497                 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1498                 .size = size,
1499         };
1500         int ret;
1501         union vni {
1502                 uint32_t vlan_id;
1503                 uint8_t vni[4];
1504         } id = { .vlan_id = 0, };
1505
1506         if (!((struct priv *)dev->data->dev_private)->config.l3_vxlan_en)
1507                 return rte_flow_error_set(error, ENOTSUP,
1508                                           RTE_FLOW_ERROR_TYPE_ITEM,
1509                                           item,
1510                                           "L3 VXLAN is not enabled by device"
1511                                           " parameter and/or not configured in"
1512                                           " firmware");
1513         if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1514                 return rte_flow_error_set(error, ENOTSUP,
1515                                           RTE_FLOW_ERROR_TYPE_ITEM,
1516                                           item,
1517                                           "a tunnel is already present");
1518         /*
1519          * Verify only UDPv4 is present as defined in
1520          * https://tools.ietf.org/html/rfc7348
1521          */
1522         if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L4_UDP))
1523                 return rte_flow_error_set(error, ENOTSUP,
1524                                           RTE_FLOW_ERROR_TYPE_ITEM,
1525                                           item,
1526                                           "no outer UDP layer found");
1527         if (!mask)
1528                 mask = &rte_flow_item_vxlan_gpe_mask;
1529         ret = mlx5_flow_item_acceptable
1530                 (item, (const uint8_t *)mask,
1531                  (const uint8_t *)&rte_flow_item_vxlan_gpe_mask,
1532                  sizeof(struct rte_flow_item_vxlan_gpe), error);
1533         if (ret < 0)
1534                 return ret;
1535         if (spec) {
1536                 memcpy(&id.vni[1], spec->vni, 3);
1537                 vxlan_gpe.val.tunnel_id = id.vlan_id;
1538                 memcpy(&id.vni[1], mask->vni, 3);
1539                 vxlan_gpe.mask.tunnel_id = id.vlan_id;
1540                 if (spec->protocol)
1541                         return rte_flow_error_set
1542                                 (error, EINVAL,
1543                                  RTE_FLOW_ERROR_TYPE_ITEM,
1544                                  item,
1545                                  "VxLAN-GPE protocol not supported");
1546                 /* Remove unwanted bits from values. */
1547                 vxlan_gpe.val.tunnel_id &= vxlan_gpe.mask.tunnel_id;
1548         }
1549         /*
1550          * Tunnel id 0 is equivalent as not adding a VXLAN layer, if only this
1551          * layer is defined in the Verbs specification it is interpreted as
1552          * wildcard and all packets will match this rule, if it follows a full
1553          * stack layer (ex: eth / ipv4 / udp), all packets matching the layers
1554          * before will also match this rule.  To avoid such situation, VNI 0
1555          * is currently refused.
1556          */
1557         if (!vxlan_gpe.val.tunnel_id)
1558                 return rte_flow_error_set(error, EINVAL,
1559                                           RTE_FLOW_ERROR_TYPE_ITEM,
1560                                           item,
1561                                           "VXLAN-GPE vni cannot be 0");
1562         if (!(flow->layers & MLX5_FLOW_LAYER_OUTER))
1563                 return rte_flow_error_set(error, EINVAL,
1564                                           RTE_FLOW_ERROR_TYPE_ITEM,
1565                                           item,
1566                                           "VXLAN-GPE tunnel must be fully"
1567                                           " defined");
1568         if (size <= flow_size) {
1569                 mlx5_flow_spec_verbs_add(flow, &vxlan_gpe, size);
1570                 flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1571         }
1572         flow->layers |= MLX5_FLOW_LAYER_VXLAN_GPE;
1573         return size;
1574 }
1575
1576 /**
1577  * Update the protocol in Verbs IPv4/IPv6 spec.
1578  *
1579  * @param[in, out] attr
1580  *   Pointer to Verbs attributes structure.
1581  * @param[in] search
1582  *   Specification type to search in order to update the IP protocol.
1583  * @param[in] protocol
1584  *   Protocol value to set if none is present in the specification.
1585  */
1586 static void
1587 mlx5_flow_item_gre_ip_protocol_update(struct ibv_flow_attr *attr,
1588                                       enum ibv_flow_spec_type search,
1589                                       uint8_t protocol)
1590 {
1591         unsigned int i;
1592         struct ibv_spec_header *hdr = (struct ibv_spec_header *)
1593                 ((uint8_t *)attr + sizeof(struct ibv_flow_attr));
1594
1595         if (!attr)
1596                 return;
1597         for (i = 0; i != attr->num_of_specs; ++i) {
1598                 if (hdr->type == search) {
1599                         union {
1600                                 struct ibv_flow_spec_ipv4_ext *ipv4;
1601                                 struct ibv_flow_spec_ipv6 *ipv6;
1602                         } ip;
1603
1604                         switch (search) {
1605                         case IBV_FLOW_SPEC_IPV4_EXT:
1606                                 ip.ipv4 = (struct ibv_flow_spec_ipv4_ext *)hdr;
1607                                 if (!ip.ipv4->val.proto) {
1608                                         ip.ipv4->val.proto = protocol;
1609                                         ip.ipv4->mask.proto = 0xff;
1610                                 }
1611                                 break;
1612                         case IBV_FLOW_SPEC_IPV6:
1613                                 ip.ipv6 = (struct ibv_flow_spec_ipv6 *)hdr;
1614                                 if (!ip.ipv6->val.next_hdr) {
1615                                         ip.ipv6->val.next_hdr = protocol;
1616                                         ip.ipv6->mask.next_hdr = 0xff;
1617                                 }
1618                                 break;
1619                         default:
1620                                 break;
1621                         }
1622                         break;
1623                 }
1624                 hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
1625         }
1626 }
1627
1628 /**
1629  * Convert the @p item into a Verbs specification after ensuring the NIC
1630  * will understand and process it correctly.
1631  * It will also update the previous L3 layer with the protocol value matching
1632  * the GRE.
1633  * If the necessary size for the conversion is greater than the @p flow_size,
1634  * nothing is written in @p flow, the validation is still performed.
1635  *
1636  * @param dev
1637  *   Pointer to Ethernet device.
1638  * @param[in] item
1639  *   Item specification.
1640  * @param[in, out] flow
1641  *   Pointer to flow structure.
1642  * @param[in] flow_size
1643  *   Size in bytes of the available space in @p flow, if too small, nothing is
1644  *   written.
1645  * @param[out] error
1646  *   Pointer to error structure.
1647  *
1648  * @return
1649  *   On success the number of bytes consumed/necessary, if the returned value
1650  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1651  *   otherwise another call with this returned memory size should be done.
1652  *   On error, a negative errno value is returned and rte_errno is set.
1653  */
1654 static int
1655 mlx5_flow_item_gre(const struct rte_flow_item *item,
1656                    struct rte_flow *flow, const size_t flow_size,
1657                    struct rte_flow_error *error)
1658 {
1659         struct mlx5_flow_verbs *verbs = flow->cur_verbs;
1660         const struct rte_flow_item_gre *spec = item->spec;
1661         const struct rte_flow_item_gre *mask = item->mask;
1662 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1663         unsigned int size = sizeof(struct ibv_flow_spec_gre);
1664         struct ibv_flow_spec_gre tunnel = {
1665                 .type = IBV_FLOW_SPEC_GRE,
1666                 .size = size,
1667         };
1668 #else
1669         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1670         struct ibv_flow_spec_tunnel tunnel = {
1671                 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
1672                 .size = size,
1673         };
1674 #endif
1675         int ret;
1676
1677         if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_GRE)
1678                 return rte_flow_error_set(error, ENOTSUP,
1679                                           RTE_FLOW_ERROR_TYPE_ITEM,
1680                                           item,
1681                                           "protocol filtering not compatible"
1682                                           " with this GRE layer");
1683         if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1684                 return rte_flow_error_set(error, ENOTSUP,
1685                                           RTE_FLOW_ERROR_TYPE_ITEM,
1686                                           item,
1687                                           "a tunnel is already present");
1688         if (!(flow->layers & MLX5_FLOW_LAYER_OUTER_L3))
1689                 return rte_flow_error_set(error, ENOTSUP,
1690                                           RTE_FLOW_ERROR_TYPE_ITEM,
1691                                           item,
1692                                           "L3 Layer is missing");
1693         if (!mask)
1694                 mask = &rte_flow_item_gre_mask;
1695         ret = mlx5_flow_item_acceptable
1696                 (item, (const uint8_t *)mask,
1697                  (const uint8_t *)&rte_flow_item_gre_mask,
1698                  sizeof(struct rte_flow_item_gre), error);
1699         if (ret < 0)
1700                 return ret;
1701 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1702         if (spec) {
1703                 tunnel.val.c_ks_res0_ver = spec->c_rsvd0_ver;
1704                 tunnel.val.protocol = spec->protocol;
1705                 tunnel.mask.c_ks_res0_ver = mask->c_rsvd0_ver;
1706                 tunnel.mask.protocol = mask->protocol;
1707                 /* Remove unwanted bits from values. */
1708                 tunnel.val.c_ks_res0_ver &= tunnel.mask.c_ks_res0_ver;
1709                 tunnel.val.protocol &= tunnel.mask.protocol;
1710                 tunnel.val.key &= tunnel.mask.key;
1711         }
1712 #else
1713         if (spec && (spec->protocol & mask->protocol))
1714                 return rte_flow_error_set(error, ENOTSUP,
1715                                           RTE_FLOW_ERROR_TYPE_ITEM,
1716                                           item,
1717                                           "without MPLS support the"
1718                                           " specification cannot be used for"
1719                                           " filtering");
1720 #endif /* !HAVE_IBV_DEVICE_MPLS_SUPPORT */
1721         if (size <= flow_size) {
1722                 if (flow->layers & MLX5_FLOW_LAYER_OUTER_L3_IPV4)
1723                         mlx5_flow_item_gre_ip_protocol_update
1724                                 (verbs->attr, IBV_FLOW_SPEC_IPV4_EXT,
1725                                  MLX5_IP_PROTOCOL_GRE);
1726                 else
1727                         mlx5_flow_item_gre_ip_protocol_update
1728                                 (verbs->attr, IBV_FLOW_SPEC_IPV6,
1729                                  MLX5_IP_PROTOCOL_GRE);
1730                 mlx5_flow_spec_verbs_add(flow, &tunnel, size);
1731                 flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1732         }
1733         flow->layers |= MLX5_FLOW_LAYER_GRE;
1734         return size;
1735 }
1736
1737 /**
1738  * Convert the @p item into a Verbs specification after ensuring the NIC
1739  * will understand and process it correctly.
1740  * If the necessary size for the conversion is greater than the @p flow_size,
1741  * nothing is written in @p flow, the validation is still performed.
1742  *
1743  * @param[in] item
1744  *   Item specification.
1745  * @param[in, out] flow
1746  *   Pointer to flow structure.
1747  * @param[in] flow_size
1748  *   Size in bytes of the available space in @p flow, if too small, nothing is
1749  *   written.
1750  * @param[out] error
1751  *   Pointer to error structure.
1752  *
1753  * @return
1754  *   On success the number of bytes consumed/necessary, if the returned value
1755  *   is lesser or equal to @p flow_size, the @p item has fully been converted,
1756  *   otherwise another call with this returned memory size should be done.
1757  *   On error, a negative errno value is returned and rte_errno is set.
1758  */
1759 static int
1760 mlx5_flow_item_mpls(const struct rte_flow_item *item __rte_unused,
1761                     struct rte_flow *flow __rte_unused,
1762                     const size_t flow_size __rte_unused,
1763                     struct rte_flow_error *error)
1764 {
1765 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1766         const struct rte_flow_item_mpls *spec = item->spec;
1767         const struct rte_flow_item_mpls *mask = item->mask;
1768         unsigned int size = sizeof(struct ibv_flow_spec_mpls);
1769         struct ibv_flow_spec_mpls mpls = {
1770                 .type = IBV_FLOW_SPEC_MPLS,
1771                 .size = size,
1772         };
1773         int ret;
1774
1775         if (flow->l3_protocol_en && flow->l3_protocol != MLX5_IP_PROTOCOL_MPLS)
1776                 return rte_flow_error_set(error, ENOTSUP,
1777                                           RTE_FLOW_ERROR_TYPE_ITEM,
1778                                           item,
1779                                           "protocol filtering not compatible"
1780                                           " with MPLS layer");
1781         if (flow->layers & MLX5_FLOW_LAYER_TUNNEL)
1782                 return rte_flow_error_set(error, ENOTSUP,
1783                                           RTE_FLOW_ERROR_TYPE_ITEM,
1784                                           item,
1785                                           "a tunnel is already"
1786                                           " present");
1787         if (!mask)
1788                 mask = &rte_flow_item_mpls_mask;
1789         ret = mlx5_flow_item_acceptable
1790                 (item, (const uint8_t *)mask,
1791                  (const uint8_t *)&rte_flow_item_mpls_mask,
1792                  sizeof(struct rte_flow_item_mpls), error);
1793         if (ret < 0)
1794                 return ret;
1795         if (spec) {
1796                 memcpy(&mpls.val.label, spec, sizeof(mpls.val.label));
1797                 memcpy(&mpls.mask.label, mask, sizeof(mpls.mask.label));
1798                 /* Remove unwanted bits from values.  */
1799                 mpls.val.label &= mpls.mask.label;
1800         }
1801         if (size <= flow_size) {
1802                 mlx5_flow_spec_verbs_add(flow, &mpls, size);
1803                 flow->cur_verbs->attr->priority = MLX5_PRIORITY_MAP_L2;
1804         }
1805         flow->layers |= MLX5_FLOW_LAYER_MPLS;
1806         return size;
1807 #endif /* !HAVE_IBV_DEVICE_MPLS_SUPPORT */
1808         return rte_flow_error_set(error, ENOTSUP,
1809                                   RTE_FLOW_ERROR_TYPE_ITEM,
1810                                   item,
1811                                   "MPLS is not supported by Verbs, please"
1812                                   " update.");
1813 }
1814
1815 /**
1816  * Convert the @p pattern into a Verbs specifications after ensuring the NIC
1817  * will understand and process it correctly.
1818  * The conversion is performed item per item, each of them is written into
1819  * the @p flow if its size is lesser or equal to @p flow_size.
1820  * Validation and memory consumption computation are still performed until the
1821  * end of @p pattern, unless an error is encountered.
1822  *
1823  * @param[in] pattern
1824  *   Flow pattern.
1825  * @param[in, out] flow
1826  *   Pointer to the rte_flow structure.
1827  * @param[in] flow_size
1828  *   Size in bytes of the available space in @p flow, if too small some
1829  *   garbage may be present.
1830  * @param[out] error
1831  *   Pointer to error structure.
1832  *
1833  * @return
1834  *   On success the number of bytes consumed/necessary, if the returned value
1835  *   is lesser or equal to @p flow_size, the @pattern  has fully been
1836  *   converted, otherwise another call with this returned memory size should
1837  *   be done.
1838  *   On error, a negative errno value is returned and rte_errno is set.
1839  */
1840 static int
1841 mlx5_flow_items(struct rte_eth_dev *dev,
1842                 const struct rte_flow_item pattern[],
1843                 struct rte_flow *flow, const size_t flow_size,
1844                 struct rte_flow_error *error)
1845 {
1846         int remain = flow_size;
1847         size_t size = 0;
1848
1849         for (; pattern->type != RTE_FLOW_ITEM_TYPE_END; pattern++) {
1850                 int ret = 0;
1851
1852                 switch (pattern->type) {
1853                 case RTE_FLOW_ITEM_TYPE_VOID:
1854                         break;
1855                 case RTE_FLOW_ITEM_TYPE_ETH:
1856                         ret = mlx5_flow_item_eth(pattern, flow, remain, error);
1857                         break;
1858                 case RTE_FLOW_ITEM_TYPE_VLAN:
1859                         ret = mlx5_flow_item_vlan(pattern, flow, remain, error);
1860                         break;
1861                 case RTE_FLOW_ITEM_TYPE_IPV4:
1862                         ret = mlx5_flow_item_ipv4(pattern, flow, remain, error);
1863                         break;
1864                 case RTE_FLOW_ITEM_TYPE_IPV6:
1865                         ret = mlx5_flow_item_ipv6(pattern, flow, remain, error);
1866                         break;
1867                 case RTE_FLOW_ITEM_TYPE_UDP:
1868                         ret = mlx5_flow_item_udp(pattern, flow, remain, error);
1869                         break;
1870                 case RTE_FLOW_ITEM_TYPE_TCP:
1871                         ret = mlx5_flow_item_tcp(pattern, flow, remain, error);
1872                         break;
1873                 case RTE_FLOW_ITEM_TYPE_VXLAN:
1874                         ret = mlx5_flow_item_vxlan(pattern, flow, remain,
1875                                                    error);
1876                         break;
1877                 case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1878                         ret = mlx5_flow_item_vxlan_gpe(dev, pattern, flow,
1879                                                        remain, error);
1880                         break;
1881                 case RTE_FLOW_ITEM_TYPE_GRE:
1882                         ret = mlx5_flow_item_gre(pattern, flow, remain, error);
1883                         break;
1884                 case RTE_FLOW_ITEM_TYPE_MPLS:
1885                         ret = mlx5_flow_item_mpls(pattern, flow, remain, error);
1886                         break;
1887                 default:
1888                         return rte_flow_error_set(error, ENOTSUP,
1889                                                   RTE_FLOW_ERROR_TYPE_ITEM,
1890                                                   pattern,
1891                                                   "item not supported");
1892                 }
1893                 if (ret < 0)
1894                         return ret;
1895                 if (remain > ret)
1896                         remain -= ret;
1897                 else
1898                         remain = 0;
1899                 size += ret;
1900         }
1901         if (!flow->layers) {
1902                 const struct rte_flow_item item = {
1903                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1904                 };
1905
1906                 return mlx5_flow_item_eth(&item, flow, flow_size, error);
1907         }
1908         return size;
1909 }
1910
1911 /**
1912  * Convert the @p action into a Verbs specification after ensuring the NIC
1913  * will understand and process it correctly.
1914  * If the necessary size for the conversion is greater than the @p flow_size,
1915  * nothing is written in @p flow, the validation is still performed.
1916  *
1917  * @param[in] action
1918  *   Action configuration.
1919  * @param[in, out] flow
1920  *   Pointer to flow structure.
1921  * @param[in] flow_size
1922  *   Size in bytes of the available space in @p flow, if too small, nothing is
1923  *   written.
1924  * @param[out] error
1925  *   Pointer to error structure.
1926  *
1927  * @return
1928  *   On success the number of bytes consumed/necessary, if the returned value
1929  *   is lesser or equal to @p flow_size, the @p action has fully been
1930  *   converted, otherwise another call with this returned memory size should
1931  *   be done.
1932  *   On error, a negative errno value is returned and rte_errno is set.
1933  */
1934 static int
1935 mlx5_flow_action_drop(const struct rte_flow_action *action,
1936                       struct rte_flow *flow, const size_t flow_size,
1937                       struct rte_flow_error *error)
1938 {
1939         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1940         struct ibv_flow_spec_action_drop drop = {
1941                         .type = IBV_FLOW_SPEC_ACTION_DROP,
1942                         .size = size,
1943         };
1944
1945         if (flow->fate)
1946                 return rte_flow_error_set(error, ENOTSUP,
1947                                           RTE_FLOW_ERROR_TYPE_ACTION,
1948                                           action,
1949                                           "multiple fate actions are not"
1950                                           " supported");
1951         if (flow->modifier & (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK))
1952                 return rte_flow_error_set(error, ENOTSUP,
1953                                           RTE_FLOW_ERROR_TYPE_ACTION,
1954                                           action,
1955                                           "drop is not compatible with"
1956                                           " flag/mark action");
1957         if (size < flow_size)
1958                 mlx5_flow_spec_verbs_add(flow, &drop, size);
1959         flow->fate |= MLX5_FLOW_FATE_DROP;
1960         return size;
1961 }
1962
1963 /**
1964  * Convert the @p action into @p flow after ensuring the NIC will understand
1965  * and process it correctly.
1966  *
1967  * @param[in] dev
1968  *   Pointer to Ethernet device structure.
1969  * @param[in] action
1970  *   Action configuration.
1971  * @param[in, out] flow
1972  *   Pointer to flow structure.
1973  * @param[out] error
1974  *   Pointer to error structure.
1975  *
1976  * @return
1977  *   0 on success, a negative errno value otherwise and rte_errno is set.
1978  */
1979 static int
1980 mlx5_flow_action_queue(struct rte_eth_dev *dev,
1981                        const struct rte_flow_action *action,
1982                        struct rte_flow *flow,
1983                        struct rte_flow_error *error)
1984 {
1985         struct priv *priv = dev->data->dev_private;
1986         const struct rte_flow_action_queue *queue = action->conf;
1987
1988         if (flow->fate)
1989                 return rte_flow_error_set(error, ENOTSUP,
1990                                           RTE_FLOW_ERROR_TYPE_ACTION,
1991                                           action,
1992                                           "multiple fate actions are not"
1993                                           " supported");
1994         if (queue->index >= priv->rxqs_n)
1995                 return rte_flow_error_set(error, EINVAL,
1996                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
1997                                           &queue->index,
1998                                           "queue index out of range");
1999         if (!(*priv->rxqs)[queue->index])
2000                 return rte_flow_error_set(error, EINVAL,
2001                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2002                                           &queue->index,
2003                                           "queue is not configured");
2004         if (flow->queue)
2005                 (*flow->queue)[0] = queue->index;
2006         flow->rss.queue_num = 1;
2007         flow->fate |= MLX5_FLOW_FATE_QUEUE;
2008         return 0;
2009 }
2010
2011 /**
2012  * Ensure the @p action will be understood and used correctly by the  NIC.
2013  *
2014  * @param dev
2015  *   Pointer to Ethernet device structure.
2016  * @param action[in]
2017  *   Pointer to flow actions array.
2018  * @param flow[in, out]
2019  *   Pointer to the rte_flow structure.
2020  * @param error[in, out]
2021  *   Pointer to error structure.
2022  *
2023  * @return
2024  *   On success @p flow->queue array and @p flow->rss are filled and valid.
2025  *   On error, a negative errno value is returned and rte_errno is set.
2026  */
2027 static int
2028 mlx5_flow_action_rss(struct rte_eth_dev *dev,
2029                      const struct rte_flow_action *action,
2030                      struct rte_flow *flow,
2031                      struct rte_flow_error *error)
2032 {
2033         struct priv *priv = dev->data->dev_private;
2034         const struct rte_flow_action_rss *rss = action->conf;
2035         unsigned int i;
2036
2037         if (flow->fate)
2038                 return rte_flow_error_set(error, ENOTSUP,
2039                                           RTE_FLOW_ERROR_TYPE_ACTION,
2040                                           action,
2041                                           "multiple fate actions are not"
2042                                           " supported");
2043         if (rss->func != RTE_ETH_HASH_FUNCTION_DEFAULT &&
2044             rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ)
2045                 return rte_flow_error_set(error, ENOTSUP,
2046                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2047                                           &rss->func,
2048                                           "RSS hash function not supported");
2049 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2050         if (rss->level > 2)
2051 #else
2052         if (rss->level > 1)
2053 #endif
2054                 return rte_flow_error_set(error, ENOTSUP,
2055                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2056                                           &rss->level,
2057                                           "tunnel RSS is not supported");
2058         if (rss->key_len < MLX5_RSS_HASH_KEY_LEN)
2059                 return rte_flow_error_set(error, ENOTSUP,
2060                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2061                                           &rss->key_len,
2062                                           "RSS hash key too small");
2063         if (rss->key_len > MLX5_RSS_HASH_KEY_LEN)
2064                 return rte_flow_error_set(error, ENOTSUP,
2065                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2066                                           &rss->key_len,
2067                                           "RSS hash key too large");
2068         if (rss->queue_num > priv->config.ind_table_max_size)
2069                 return rte_flow_error_set(error, ENOTSUP,
2070                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2071                                           &rss->queue_num,
2072                                           "number of queues too large");
2073         if (rss->types & MLX5_RSS_HF_MASK)
2074                 return rte_flow_error_set(error, ENOTSUP,
2075                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2076                                           &rss->types,
2077                                           "some RSS protocols are not"
2078                                           " supported");
2079         for (i = 0; i != rss->queue_num; ++i) {
2080                 if (!(*priv->rxqs)[rss->queue[i]])
2081                         return rte_flow_error_set
2082                                 (error, EINVAL,
2083                                  RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2084                                  &rss->queue[i],
2085                                  "queue is not configured");
2086         }
2087         if (flow->queue)
2088                 memcpy((*flow->queue), rss->queue,
2089                        rss->queue_num * sizeof(uint16_t));
2090         flow->rss.queue_num = rss->queue_num;
2091         memcpy(flow->key, rss->key, MLX5_RSS_HASH_KEY_LEN);
2092         flow->rss.types = rss->types;
2093         flow->rss.level = rss->level;
2094         flow->fate |= MLX5_FLOW_FATE_RSS;
2095         return 0;
2096 }
2097
2098 /**
2099  * Convert the @p action into a Verbs specification after ensuring the NIC
2100  * will understand and process it correctly.
2101  * If the necessary size for the conversion is greater than the @p flow_size,
2102  * nothing is written in @p flow, the validation is still performed.
2103  *
2104  * @param[in] action
2105  *   Action configuration.
2106  * @param[in, out] flow
2107  *   Pointer to flow structure.
2108  * @param[in] flow_size
2109  *   Size in bytes of the available space in @p flow, if too small, nothing is
2110  *   written.
2111  * @param[out] error
2112  *   Pointer to error structure.
2113  *
2114  * @return
2115  *   On success the number of bytes consumed/necessary, if the returned value
2116  *   is lesser or equal to @p flow_size, the @p action has fully been
2117  *   converted, otherwise another call with this returned memory size should
2118  *   be done.
2119  *   On error, a negative errno value is returned and rte_errno is set.
2120  */
2121 static int
2122 mlx5_flow_action_flag(const struct rte_flow_action *action,
2123                       struct rte_flow *flow, const size_t flow_size,
2124                       struct rte_flow_error *error)
2125 {
2126         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
2127         struct ibv_flow_spec_action_tag tag = {
2128                 .type = IBV_FLOW_SPEC_ACTION_TAG,
2129                 .size = size,
2130                 .tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
2131         };
2132         struct mlx5_flow_verbs *verbs = flow->cur_verbs;
2133
2134         if (flow->modifier & MLX5_FLOW_MOD_FLAG)
2135                 return rte_flow_error_set(error, ENOTSUP,
2136                                           RTE_FLOW_ERROR_TYPE_ACTION,
2137                                           action,
2138                                           "flag action already present");
2139         if (flow->fate & MLX5_FLOW_FATE_DROP)
2140                 return rte_flow_error_set(error, ENOTSUP,
2141                                           RTE_FLOW_ERROR_TYPE_ACTION,
2142                                           action,
2143                                           "flag is not compatible with drop"
2144                                           " action");
2145         if (flow->modifier & MLX5_FLOW_MOD_MARK)
2146                 size = 0;
2147         else if (size <= flow_size && verbs)
2148                 mlx5_flow_spec_verbs_add(flow, &tag, size);
2149         flow->modifier |= MLX5_FLOW_MOD_FLAG;
2150         return size;
2151 }
2152
2153 /**
2154  * Update verbs specification to modify the flag to mark.
2155  *
2156  * @param[in, out] verbs
2157  *   Pointer to the mlx5_flow_verbs structure.
2158  * @param[in] mark_id
2159  *   Mark identifier to replace the flag.
2160  */
2161 static void
2162 mlx5_flow_verbs_mark_update(struct mlx5_flow_verbs *verbs, uint32_t mark_id)
2163 {
2164         struct ibv_spec_header *hdr;
2165         int i;
2166
2167         if (!verbs)
2168                 return;
2169         /* Update Verbs specification. */
2170         hdr = (struct ibv_spec_header *)verbs->specs;
2171         if (!hdr)
2172                 return;
2173         for (i = 0; i != verbs->attr->num_of_specs; ++i) {
2174                 if (hdr->type == IBV_FLOW_SPEC_ACTION_TAG) {
2175                         struct ibv_flow_spec_action_tag *t =
2176                                 (struct ibv_flow_spec_action_tag *)hdr;
2177
2178                         t->tag_id = mlx5_flow_mark_set(mark_id);
2179                 }
2180                 hdr = (struct ibv_spec_header *)((uintptr_t)hdr + hdr->size);
2181         }
2182 }
2183
2184 /**
2185  * Convert the @p action into @p flow (or by updating the already present
2186  * Flag Verbs specification) after ensuring the NIC will understand and
2187  * process it correctly.
2188  * If the necessary size for the conversion is greater than the @p flow_size,
2189  * nothing is written in @p flow, the validation is still performed.
2190  *
2191  * @param[in] action
2192  *   Action configuration.
2193  * @param[in, out] flow
2194  *   Pointer to flow structure.
2195  * @param[in] flow_size
2196  *   Size in bytes of the available space in @p flow, if too small, nothing is
2197  *   written.
2198  * @param[out] error
2199  *   Pointer to error structure.
2200  *
2201  * @return
2202  *   On success the number of bytes consumed/necessary, if the returned value
2203  *   is lesser or equal to @p flow_size, the @p action has fully been
2204  *   converted, otherwise another call with this returned memory size should
2205  *   be done.
2206  *   On error, a negative errno value is returned and rte_errno is set.
2207  */
2208 static int
2209 mlx5_flow_action_mark(const struct rte_flow_action *action,
2210                       struct rte_flow *flow, const size_t flow_size,
2211                       struct rte_flow_error *error)
2212 {
2213         const struct rte_flow_action_mark *mark = action->conf;
2214         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
2215         struct ibv_flow_spec_action_tag tag = {
2216                 .type = IBV_FLOW_SPEC_ACTION_TAG,
2217                 .size = size,
2218         };
2219         struct mlx5_flow_verbs *verbs = flow->cur_verbs;
2220
2221         if (!mark)
2222                 return rte_flow_error_set(error, EINVAL,
2223                                           RTE_FLOW_ERROR_TYPE_ACTION,
2224                                           action,
2225                                           "configuration cannot be null");
2226         if (mark->id >= MLX5_FLOW_MARK_MAX)
2227                 return rte_flow_error_set(error, EINVAL,
2228                                           RTE_FLOW_ERROR_TYPE_ACTION_CONF,
2229                                           &mark->id,
2230                                           "mark id must in 0 <= id < "
2231                                           RTE_STR(MLX5_FLOW_MARK_MAX));
2232         if (flow->modifier & MLX5_FLOW_MOD_MARK)
2233                 return rte_flow_error_set(error, ENOTSUP,
2234                                           RTE_FLOW_ERROR_TYPE_ACTION,
2235                                           action,
2236                                           "mark action already present");
2237         if (flow->fate & MLX5_FLOW_FATE_DROP)
2238                 return rte_flow_error_set(error, ENOTSUP,
2239                                           RTE_FLOW_ERROR_TYPE_ACTION,
2240                                           action,
2241                                           "mark is not compatible with drop"
2242                                           " action");
2243         if (flow->modifier & MLX5_FLOW_MOD_FLAG) {
2244                 mlx5_flow_verbs_mark_update(verbs, mark->id);
2245                 size = 0;
2246         } else if (size <= flow_size) {
2247                 tag.tag_id = mlx5_flow_mark_set(mark->id);
2248                 mlx5_flow_spec_verbs_add(flow, &tag, size);
2249         }
2250         flow->modifier |= MLX5_FLOW_MOD_MARK;
2251         return size;
2252 }
2253
2254 /**
2255  * Convert the @p action into a Verbs specification after ensuring the NIC
2256  * will understand and process it correctly.
2257  * If the necessary size for the conversion is greater than the @p flow_size,
2258  * nothing is written in @p flow, the validation is still performed.
2259  *
2260  * @param action[in]
2261  *   Action configuration.
2262  * @param flow[in, out]
2263  *   Pointer to flow structure.
2264  * @param flow_size[in]
2265  *   Size in bytes of the available space in @p flow, if too small, nothing is
2266  *   written.
2267  * @param error[int, out]
2268  *   Pointer to error structure.
2269  *
2270  * @return
2271  *   On success the number of bytes consumed/necessary, if the returned value
2272  *   is lesser or equal to @p flow_size, the @p action has fully been
2273  *   converted, otherwise another call with this returned memory size should
2274  *   be done.
2275  *   On error, a negative errno value is returned and rte_errno is set.
2276  */
2277 static int
2278 mlx5_flow_action_count(struct rte_eth_dev *dev,
2279                        const struct rte_flow_action *action,
2280                        struct rte_flow *flow,
2281                        const size_t flow_size __rte_unused,
2282                        struct rte_flow_error *error)
2283 {
2284         const struct rte_flow_action_count *count = action->conf;
2285 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2286         unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
2287         struct ibv_flow_spec_counter_action counter = {
2288                 .type = IBV_FLOW_SPEC_ACTION_COUNT,
2289                 .size = size,
2290         };
2291 #endif
2292
2293         if (!flow->counter) {
2294                 flow->counter = mlx5_flow_counter_new(dev, count->shared,
2295                                                       count->id);
2296                 if (!flow->counter)
2297                         return rte_flow_error_set(error, ENOTSUP,
2298                                                   RTE_FLOW_ERROR_TYPE_ACTION,
2299                                                   action,
2300                                                   "cannot get counter"
2301                                                   " context.");
2302         }
2303         if (!((struct priv *)dev->data->dev_private)->config.flow_counter_en)
2304                 return rte_flow_error_set(error, ENOTSUP,
2305                                           RTE_FLOW_ERROR_TYPE_ACTION,
2306                                           action,
2307                                           "flow counters are not supported.");
2308         flow->modifier |= MLX5_FLOW_MOD_COUNT;
2309 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2310         counter.counter_set_handle = flow->counter->cs->handle;
2311         if (size <= flow_size)
2312                 mlx5_flow_spec_verbs_add(flow, &counter, size);
2313         return size;
2314 #endif
2315         return 0;
2316 }
2317
2318 /**
2319  * Convert the @p action into @p flow after ensuring the NIC will understand
2320  * and process it correctly.
2321  * The conversion is performed action per action, each of them is written into
2322  * the @p flow if its size is lesser or equal to @p flow_size.
2323  * Validation and memory consumption computation are still performed until the
2324  * end of @p action, unless an error is encountered.
2325  *
2326  * @param[in] dev
2327  *   Pointer to Ethernet device structure.
2328  * @param[in] actions
2329  *   Pointer to flow actions array.
2330  * @param[in, out] flow
2331  *   Pointer to the rte_flow structure.
2332  * @param[in] flow_size
2333  *   Size in bytes of the available space in @p flow, if too small some
2334  *   garbage may be present.
2335  * @param[out] error
2336  *   Pointer to error structure.
2337  *
2338  * @return
2339  *   On success the number of bytes consumed/necessary, if the returned value
2340  *   is lesser or equal to @p flow_size, the @p actions has fully been
2341  *   converted, otherwise another call with this returned memory size should
2342  *   be done.
2343  *   On error, a negative errno value is returned and rte_errno is set.
2344  */
2345 static int
2346 mlx5_flow_actions(struct rte_eth_dev *dev,
2347                   const struct rte_flow_action actions[],
2348                   struct rte_flow *flow, const size_t flow_size,
2349                   struct rte_flow_error *error)
2350 {
2351         size_t size = 0;
2352         int remain = flow_size;
2353         int ret = 0;
2354
2355         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
2356                 switch (actions->type) {
2357                 case RTE_FLOW_ACTION_TYPE_VOID:
2358                         break;
2359                 case RTE_FLOW_ACTION_TYPE_FLAG:
2360                         ret = mlx5_flow_action_flag(actions, flow, remain,
2361                                                     error);
2362                         break;
2363                 case RTE_FLOW_ACTION_TYPE_MARK:
2364                         ret = mlx5_flow_action_mark(actions, flow, remain,
2365                                                     error);
2366                         break;
2367                 case RTE_FLOW_ACTION_TYPE_DROP:
2368                         ret = mlx5_flow_action_drop(actions, flow, remain,
2369                                                     error);
2370                         break;
2371                 case RTE_FLOW_ACTION_TYPE_QUEUE:
2372                         ret = mlx5_flow_action_queue(dev, actions, flow, error);
2373                         break;
2374                 case RTE_FLOW_ACTION_TYPE_RSS:
2375                         ret = mlx5_flow_action_rss(dev, actions, flow, error);
2376                         break;
2377                 case RTE_FLOW_ACTION_TYPE_COUNT:
2378                         ret = mlx5_flow_action_count(dev, actions, flow, remain,
2379                                                      error);
2380                         break;
2381                 default:
2382                         return rte_flow_error_set(error, ENOTSUP,
2383                                                   RTE_FLOW_ERROR_TYPE_ACTION,
2384                                                   actions,
2385                                                   "action not supported");
2386                 }
2387                 if (ret < 0)
2388                         return ret;
2389                 if (remain > ret)
2390                         remain -= ret;
2391                 else
2392                         remain = 0;
2393                 size += ret;
2394         }
2395         if (!flow->fate)
2396                 return rte_flow_error_set(error, ENOTSUP,
2397                                           RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2398                                           NULL,
2399                                           "no fate action found");
2400         return size;
2401 }
2402
2403 /**
2404  * Validate flow rule and fill flow structure accordingly.
2405  *
2406  * @param dev
2407  *   Pointer to Ethernet device.
2408  * @param[out] flow
2409  *   Pointer to flow structure.
2410  * @param flow_size
2411  *   Size of allocated space for @p flow.
2412  * @param[in] attr
2413  *   Flow rule attributes.
2414  * @param[in] pattern
2415  *   Pattern specification (list terminated by the END pattern item).
2416  * @param[in] actions
2417  *   Associated actions (list terminated by the END action).
2418  * @param[out] error
2419  *   Perform verbose error reporting if not NULL.
2420  *
2421  * @return
2422  *   A positive value representing the size of the flow object in bytes
2423  *   regardless of @p flow_size on success, a negative errno value otherwise
2424  *   and rte_errno is set.
2425  */
2426 static int
2427 mlx5_flow_merge_switch(struct rte_eth_dev *dev,
2428                        struct rte_flow *flow,
2429                        size_t flow_size,
2430                        const struct rte_flow_attr *attr,
2431                        const struct rte_flow_item pattern[],
2432                        const struct rte_flow_action actions[],
2433                        struct rte_flow_error *error)
2434 {
2435         unsigned int n = mlx5_dev_to_port_id(dev->device, NULL, 0);
2436         uint16_t port_id[!n + n];
2437         struct mlx5_nl_flow_ptoi ptoi[!n + n + 1];
2438         size_t off = RTE_ALIGN_CEIL(sizeof(*flow), alignof(max_align_t));
2439         unsigned int i;
2440         unsigned int own = 0;
2441         int ret;
2442
2443         /* At least one port is needed when no switch domain is present. */
2444         if (!n) {
2445                 n = 1;
2446                 port_id[0] = dev->data->port_id;
2447         } else {
2448                 n = RTE_MIN(mlx5_dev_to_port_id(dev->device, port_id, n), n);
2449         }
2450         for (i = 0; i != n; ++i) {
2451                 struct rte_eth_dev_info dev_info;
2452
2453                 rte_eth_dev_info_get(port_id[i], &dev_info);
2454                 if (port_id[i] == dev->data->port_id)
2455                         own = i;
2456                 ptoi[i].port_id = port_id[i];
2457                 ptoi[i].ifindex = dev_info.if_index;
2458         }
2459         /* Ensure first entry of ptoi[] is the current device. */
2460         if (own) {
2461                 ptoi[n] = ptoi[0];
2462                 ptoi[0] = ptoi[own];
2463                 ptoi[own] = ptoi[n];
2464         }
2465         /* An entry with zero ifindex terminates ptoi[]. */
2466         ptoi[n].port_id = 0;
2467         ptoi[n].ifindex = 0;
2468         if (flow_size < off)
2469                 flow_size = 0;
2470         ret = mlx5_nl_flow_transpose((uint8_t *)flow + off,
2471                                      flow_size ? flow_size - off : 0,
2472                                      ptoi, attr, pattern, actions, error);
2473         if (ret < 0)
2474                 return ret;
2475         if (flow_size) {
2476                 *flow = (struct rte_flow){
2477                         .attributes = *attr,
2478                         .nl_flow = (uint8_t *)flow + off,
2479                 };
2480                 /*
2481                  * Generate a reasonably unique handle based on the address
2482                  * of the target buffer.
2483                  *
2484                  * This is straightforward on 32-bit systems where the flow
2485                  * pointer can be used directly. Otherwise, its least
2486                  * significant part is taken after shifting it by the
2487                  * previous power of two of the pointed buffer size.
2488                  */
2489                 if (sizeof(flow) <= 4)
2490                         mlx5_nl_flow_brand(flow->nl_flow, (uintptr_t)flow);
2491                 else
2492                         mlx5_nl_flow_brand
2493                                 (flow->nl_flow,
2494                                  (uintptr_t)flow >>
2495                                  rte_log2_u32(rte_align32prevpow2(flow_size)));
2496         }
2497         return off + ret;
2498 }
2499
2500 static unsigned int
2501 mlx5_find_graph_root(const struct rte_flow_item pattern[], uint32_t rss_level)
2502 {
2503         const struct rte_flow_item *item;
2504         unsigned int has_vlan = 0;
2505
2506         for (item = pattern; item->type != RTE_FLOW_ITEM_TYPE_END; item++) {
2507                 if (item->type == RTE_FLOW_ITEM_TYPE_VLAN) {
2508                         has_vlan = 1;
2509                         break;
2510                 }
2511         }
2512         if (has_vlan)
2513                 return rss_level < 2 ? MLX5_EXPANSION_ROOT_ETH_VLAN :
2514                                        MLX5_EXPANSION_ROOT_OUTER_ETH_VLAN;
2515         return rss_level < 2 ? MLX5_EXPANSION_ROOT :
2516                                MLX5_EXPANSION_ROOT_OUTER;
2517 }
2518
2519 /**
2520  * Convert the @p attributes, @p pattern, @p action, into an flow for the NIC
2521  * after ensuring the NIC will understand and process it correctly.
2522  * The conversion is only performed item/action per item/action, each of
2523  * them is written into the @p flow if its size is lesser or equal to @p
2524  * flow_size.
2525  * Validation and memory consumption computation are still performed until the
2526  * end, unless an error is encountered.
2527  *
2528  * @param[in] dev
2529  *   Pointer to Ethernet device.
2530  * @param[in, out] flow
2531  *   Pointer to flow structure.
2532  * @param[in] flow_size
2533  *   Size in bytes of the available space in @p flow, if too small some
2534  *   garbage may be present.
2535  * @param[in] attributes
2536  *   Flow rule attributes.
2537  * @param[in] pattern
2538  *   Pattern specification (list terminated by the END pattern item).
2539  * @param[in] actions
2540  *   Associated actions (list terminated by the END action).
2541  * @param[out] error
2542  *   Perform verbose error reporting if not NULL.
2543  *
2544  * @return
2545  *   On success the number of bytes consumed/necessary, if the returned value
2546  *   is lesser or equal to @p flow_size, the flow has fully been converted and
2547  *   can be applied, otherwise another call with this returned memory size
2548  *   should be done.
2549  *   On error, a negative errno value is returned and rte_errno is set.
2550  */
2551 static int
2552 mlx5_flow_merge(struct rte_eth_dev *dev, struct rte_flow *flow,
2553                 const size_t flow_size,
2554                 const struct rte_flow_attr *attributes,
2555                 const struct rte_flow_item pattern[],
2556                 const struct rte_flow_action actions[],
2557                 struct rte_flow_error *error)
2558 {
2559         struct rte_flow local_flow = { .layers = 0, };
2560         size_t size = sizeof(*flow);
2561         union {
2562                 struct rte_flow_expand_rss buf;
2563                 uint8_t buffer[2048];
2564         } expand_buffer;
2565         struct rte_flow_expand_rss *buf = &expand_buffer.buf;
2566         struct mlx5_flow_verbs *original_verbs = NULL;
2567         size_t original_verbs_size = 0;
2568         uint32_t original_layers = 0;
2569         int expanded_pattern_idx = 0;
2570         int ret;
2571         uint32_t i;
2572
2573         if (attributes->transfer)
2574                 return mlx5_flow_merge_switch(dev, flow, flow_size,
2575                                               attributes, pattern,
2576                                               actions, error);
2577         if (size > flow_size)
2578                 flow = &local_flow;
2579         ret = mlx5_flow_attributes(dev, attributes, flow, error);
2580         if (ret < 0)
2581                 return ret;
2582         ret = mlx5_flow_actions(dev, actions, &local_flow, 0, error);
2583         if (ret < 0)
2584                 return ret;
2585         if (local_flow.rss.types) {
2586                 unsigned int graph_root;
2587
2588                 graph_root = mlx5_find_graph_root(pattern,
2589                                                   local_flow.rss.level);
2590                 ret = rte_flow_expand_rss(buf, sizeof(expand_buffer.buffer),
2591                                           pattern, local_flow.rss.types,
2592                                           mlx5_support_expansion,
2593                                           graph_root);
2594                 assert(ret > 0 &&
2595                        (unsigned int)ret < sizeof(expand_buffer.buffer));
2596         } else {
2597                 buf->entries = 1;
2598                 buf->entry[0].pattern = (void *)(uintptr_t)pattern;
2599         }
2600         size += RTE_ALIGN_CEIL(local_flow.rss.queue_num * sizeof(uint16_t),
2601                                sizeof(void *));
2602         if (size <= flow_size)
2603                 flow->queue = (void *)(flow + 1);
2604         LIST_INIT(&flow->verbs);
2605         flow->layers = 0;
2606         flow->modifier = 0;
2607         flow->fate = 0;
2608         for (i = 0; i != buf->entries; ++i) {
2609                 size_t off = size;
2610                 size_t off2;
2611
2612                 flow->layers = original_layers;
2613                 size += sizeof(struct ibv_flow_attr) +
2614                         sizeof(struct mlx5_flow_verbs);
2615                 off2 = size;
2616                 if (size < flow_size) {
2617                         flow->cur_verbs = (void *)((uintptr_t)flow + off);
2618                         flow->cur_verbs->attr = (void *)(flow->cur_verbs + 1);
2619                         flow->cur_verbs->specs =
2620                                 (void *)(flow->cur_verbs->attr + 1);
2621                 }
2622                 /* First iteration convert the pattern into Verbs. */
2623                 if (i == 0) {
2624                         /* Actions don't need to be converted several time. */
2625                         ret = mlx5_flow_actions(dev, actions, flow,
2626                                                 (size < flow_size) ?
2627                                                 flow_size - size : 0,
2628                                                 error);
2629                         if (ret < 0)
2630                                 return ret;
2631                         size += ret;
2632                 } else {
2633                         /*
2634                          * Next iteration means the pattern has already been
2635                          * converted and an expansion is necessary to match
2636                          * the user RSS request.  For that only the expanded
2637                          * items will be converted, the common part with the
2638                          * user pattern are just copied into the next buffer
2639                          * zone.
2640                          */
2641                         size += original_verbs_size;
2642                         if (size < flow_size) {
2643                                 rte_memcpy(flow->cur_verbs->attr,
2644                                            original_verbs->attr,
2645                                            original_verbs_size +
2646                                            sizeof(struct ibv_flow_attr));
2647                                 flow->cur_verbs->size = original_verbs_size;
2648                         }
2649                 }
2650                 ret = mlx5_flow_items
2651                         (dev,
2652                          (const struct rte_flow_item *)
2653                          &buf->entry[i].pattern[expanded_pattern_idx],
2654                          flow,
2655                          (size < flow_size) ? flow_size - size : 0, error);
2656                 if (ret < 0)
2657                         return ret;
2658                 size += ret;
2659                 if (size <= flow_size) {
2660                         mlx5_flow_adjust_priority(dev, flow);
2661                         LIST_INSERT_HEAD(&flow->verbs, flow->cur_verbs, next);
2662                 }
2663                 /*
2664                  * Keep a pointer of the first verbs conversion and the layers
2665                  * it has encountered.
2666                  */
2667                 if (i == 0) {
2668                         original_verbs = flow->cur_verbs;
2669                         original_verbs_size = size - off2;
2670                         original_layers = flow->layers;
2671                         /*
2672                          * move the index of the expanded pattern to the
2673                          * first item not addressed yet.
2674                          */
2675                         if (pattern->type == RTE_FLOW_ITEM_TYPE_END) {
2676                                 expanded_pattern_idx++;
2677                         } else {
2678                                 const struct rte_flow_item *item = pattern;
2679
2680                                 for (item = pattern;
2681                                      item->type != RTE_FLOW_ITEM_TYPE_END;
2682                                      ++item)
2683                                         expanded_pattern_idx++;
2684                         }
2685                 }
2686         }
2687         /* Restore the origin layers in the flow. */
2688         flow->layers = original_layers;
2689         return size;
2690 }
2691
2692 /**
2693  * Lookup and set the ptype in the data Rx part.  A single Ptype can be used,
2694  * if several tunnel rules are used on this queue, the tunnel ptype will be
2695  * cleared.
2696  *
2697  * @param rxq_ctrl
2698  *   Rx queue to update.
2699  */
2700 static void
2701 mlx5_flow_rxq_tunnel_ptype_update(struct mlx5_rxq_ctrl *rxq_ctrl)
2702 {
2703         unsigned int i;
2704         uint32_t tunnel_ptype = 0;
2705
2706         /* Look up for the ptype to use. */
2707         for (i = 0; i != MLX5_FLOW_TUNNEL; ++i) {
2708                 if (!rxq_ctrl->flow_tunnels_n[i])
2709                         continue;
2710                 if (!tunnel_ptype) {
2711                         tunnel_ptype = tunnels_info[i].ptype;
2712                 } else {
2713                         tunnel_ptype = 0;
2714                         break;
2715                 }
2716         }
2717         rxq_ctrl->rxq.tunnel = tunnel_ptype;
2718 }
2719
2720 /**
2721  * Set the Rx queue flags (Mark/Flag and Tunnel Ptypes) according to the flow.
2722  *
2723  * @param[in] dev
2724  *   Pointer to Ethernet device.
2725  * @param[in] flow
2726  *   Pointer to flow structure.
2727  */
2728 static void
2729 mlx5_flow_rxq_flags_set(struct rte_eth_dev *dev, struct rte_flow *flow)
2730 {
2731         struct priv *priv = dev->data->dev_private;
2732         const int mark = !!(flow->modifier &
2733                             (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK));
2734         const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
2735         unsigned int i;
2736
2737         for (i = 0; i != flow->rss.queue_num; ++i) {
2738                 int idx = (*flow->queue)[i];
2739                 struct mlx5_rxq_ctrl *rxq_ctrl =
2740                         container_of((*priv->rxqs)[idx],
2741                                      struct mlx5_rxq_ctrl, rxq);
2742
2743                 if (mark) {
2744                         rxq_ctrl->rxq.mark = 1;
2745                         rxq_ctrl->flow_mark_n++;
2746                 }
2747                 if (tunnel) {
2748                         unsigned int j;
2749
2750                         /* Increase the counter matching the flow. */
2751                         for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
2752                                 if ((tunnels_info[j].tunnel & flow->layers) ==
2753                                     tunnels_info[j].tunnel) {
2754                                         rxq_ctrl->flow_tunnels_n[j]++;
2755                                         break;
2756                                 }
2757                         }
2758                         mlx5_flow_rxq_tunnel_ptype_update(rxq_ctrl);
2759                 }
2760         }
2761 }
2762
2763 /**
2764  * Clear the Rx queue flags (Mark/Flag and Tunnel Ptype) associated with the
2765  * @p flow if no other flow uses it with the same kind of request.
2766  *
2767  * @param dev
2768  *   Pointer to Ethernet device.
2769  * @param[in] flow
2770  *   Pointer to the flow.
2771  */
2772 static void
2773 mlx5_flow_rxq_flags_trim(struct rte_eth_dev *dev, struct rte_flow *flow)
2774 {
2775         struct priv *priv = dev->data->dev_private;
2776         const int mark = !!(flow->modifier &
2777                             (MLX5_FLOW_MOD_FLAG | MLX5_FLOW_MOD_MARK));
2778         const int tunnel = !!(flow->layers & MLX5_FLOW_LAYER_TUNNEL);
2779         unsigned int i;
2780
2781         assert(dev->data->dev_started);
2782         for (i = 0; i != flow->rss.queue_num; ++i) {
2783                 int idx = (*flow->queue)[i];
2784                 struct mlx5_rxq_ctrl *rxq_ctrl =
2785                         container_of((*priv->rxqs)[idx],
2786                                      struct mlx5_rxq_ctrl, rxq);
2787
2788                 if (mark) {
2789                         rxq_ctrl->flow_mark_n--;
2790                         rxq_ctrl->rxq.mark = !!rxq_ctrl->flow_mark_n;
2791                 }
2792                 if (tunnel) {
2793                         unsigned int j;
2794
2795                         /* Decrease the counter matching the flow. */
2796                         for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
2797                                 if ((tunnels_info[j].tunnel & flow->layers) ==
2798                                     tunnels_info[j].tunnel) {
2799                                         rxq_ctrl->flow_tunnels_n[j]--;
2800                                         break;
2801                                 }
2802                         }
2803                         mlx5_flow_rxq_tunnel_ptype_update(rxq_ctrl);
2804                 }
2805         }
2806 }
2807
2808 /**
2809  * Clear the Mark/Flag and Tunnel ptype information in all Rx queues.
2810  *
2811  * @param dev
2812  *   Pointer to Ethernet device.
2813  */
2814 static void
2815 mlx5_flow_rxq_flags_clear(struct rte_eth_dev *dev)
2816 {
2817         struct priv *priv = dev->data->dev_private;
2818         unsigned int i;
2819
2820         for (i = 0; i != priv->rxqs_n; ++i) {
2821                 struct mlx5_rxq_ctrl *rxq_ctrl;
2822                 unsigned int j;
2823
2824                 if (!(*priv->rxqs)[i])
2825                         continue;
2826                 rxq_ctrl = container_of((*priv->rxqs)[i],
2827                                         struct mlx5_rxq_ctrl, rxq);
2828                 rxq_ctrl->flow_mark_n = 0;
2829                 rxq_ctrl->rxq.mark = 0;
2830                 for (j = 0; j != MLX5_FLOW_TUNNEL; ++j)
2831                         rxq_ctrl->flow_tunnels_n[j] = 0;
2832                 rxq_ctrl->rxq.tunnel = 0;
2833         }
2834 }
2835
2836 /**
2837  * Validate a flow supported by the NIC.
2838  *
2839  * @see rte_flow_validate()
2840  * @see rte_flow_ops
2841  */
2842 int
2843 mlx5_flow_validate(struct rte_eth_dev *dev,
2844                    const struct rte_flow_attr *attr,
2845                    const struct rte_flow_item items[],
2846                    const struct rte_flow_action actions[],
2847                    struct rte_flow_error *error)
2848 {
2849         int ret = mlx5_flow_merge(dev, NULL, 0, attr, items, actions, error);
2850
2851         if (ret < 0)
2852                 return ret;
2853         return 0;
2854 }
2855
2856 /**
2857  * Remove the flow.
2858  *
2859  * @param[in] dev
2860  *   Pointer to Ethernet device.
2861  * @param[in, out] flow
2862  *   Pointer to flow structure.
2863  */
2864 static void
2865 mlx5_flow_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
2866 {
2867         struct priv *priv = dev->data->dev_private;
2868         struct mlx5_flow_verbs *verbs;
2869
2870         if (flow->nl_flow && priv->mnl_socket)
2871                 mlx5_nl_flow_destroy(priv->mnl_socket, flow->nl_flow, NULL);
2872         LIST_FOREACH(verbs, &flow->verbs, next) {
2873                 if (verbs->flow) {
2874                         claim_zero(mlx5_glue->destroy_flow(verbs->flow));
2875                         verbs->flow = NULL;
2876                 }
2877                 if (verbs->hrxq) {
2878                         if (flow->fate & MLX5_FLOW_FATE_DROP)
2879                                 mlx5_hrxq_drop_release(dev);
2880                         else
2881                                 mlx5_hrxq_release(dev, verbs->hrxq);
2882                         verbs->hrxq = NULL;
2883                 }
2884         }
2885         if (flow->counter) {
2886                 mlx5_flow_counter_release(flow->counter);
2887                 flow->counter = NULL;
2888         }
2889 }
2890
2891 /**
2892  * Apply the flow.
2893  *
2894  * @param[in] dev
2895  *   Pointer to Ethernet device structure.
2896  * @param[in, out] flow
2897  *   Pointer to flow structure.
2898  * @param[out] error
2899  *   Pointer to error structure.
2900  *
2901  * @return
2902  *   0 on success, a negative errno value otherwise and rte_errno is set.
2903  */
2904 static int
2905 mlx5_flow_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
2906                 struct rte_flow_error *error)
2907 {
2908         struct priv *priv = dev->data->dev_private;
2909         struct mlx5_flow_verbs *verbs;
2910         int err;
2911
2912         LIST_FOREACH(verbs, &flow->verbs, next) {
2913                 if (flow->fate & MLX5_FLOW_FATE_DROP) {
2914                         verbs->hrxq = mlx5_hrxq_drop_new(dev);
2915                         if (!verbs->hrxq) {
2916                                 rte_flow_error_set
2917                                         (error, errno,
2918                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2919                                          NULL,
2920                                          "cannot get drop hash queue");
2921                                 goto error;
2922                         }
2923                 } else {
2924                         struct mlx5_hrxq *hrxq;
2925
2926                         hrxq = mlx5_hrxq_get(dev, flow->key,
2927                                              MLX5_RSS_HASH_KEY_LEN,
2928                                              verbs->hash_fields,
2929                                              (*flow->queue),
2930                                              flow->rss.queue_num);
2931                         if (!hrxq)
2932                                 hrxq = mlx5_hrxq_new(dev, flow->key,
2933                                                      MLX5_RSS_HASH_KEY_LEN,
2934                                                      verbs->hash_fields,
2935                                                      (*flow->queue),
2936                                                      flow->rss.queue_num,
2937                                                      !!(flow->layers &
2938                                                       MLX5_FLOW_LAYER_TUNNEL));
2939                         if (!hrxq) {
2940                                 rte_flow_error_set
2941                                         (error, rte_errno,
2942                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2943                                          NULL,
2944                                          "cannot get hash queue");
2945                                 goto error;
2946                         }
2947                         verbs->hrxq = hrxq;
2948                 }
2949                 verbs->flow =
2950                         mlx5_glue->create_flow(verbs->hrxq->qp, verbs->attr);
2951                 if (!verbs->flow) {
2952                         rte_flow_error_set(error, errno,
2953                                            RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2954                                            NULL,
2955                                            "hardware refuses to create flow");
2956                         goto error;
2957                 }
2958         }
2959         if (flow->nl_flow &&
2960             priv->mnl_socket &&
2961             mlx5_nl_flow_create(priv->mnl_socket, flow->nl_flow, error))
2962                 goto error;
2963         return 0;
2964 error:
2965         err = rte_errno; /* Save rte_errno before cleanup. */
2966         LIST_FOREACH(verbs, &flow->verbs, next) {
2967                 if (verbs->hrxq) {
2968                         if (flow->fate & MLX5_FLOW_FATE_DROP)
2969                                 mlx5_hrxq_drop_release(dev);
2970                         else
2971                                 mlx5_hrxq_release(dev, verbs->hrxq);
2972                         verbs->hrxq = NULL;
2973                 }
2974         }
2975         rte_errno = err; /* Restore rte_errno. */
2976         return -rte_errno;
2977 }
2978
2979 /**
2980  * Create a flow and add it to @p list.
2981  *
2982  * @param dev
2983  *   Pointer to Ethernet device.
2984  * @param list
2985  *   Pointer to a TAILQ flow list.
2986  * @param[in] attr
2987  *   Flow rule attributes.
2988  * @param[in] items
2989  *   Pattern specification (list terminated by the END pattern item).
2990  * @param[in] actions
2991  *   Associated actions (list terminated by the END action).
2992  * @param[out] error
2993  *   Perform verbose error reporting if not NULL.
2994  *
2995  * @return
2996  *   A flow on success, NULL otherwise and rte_errno is set.
2997  */
2998 static struct rte_flow *
2999 mlx5_flow_list_create(struct rte_eth_dev *dev,
3000                       struct mlx5_flows *list,
3001                       const struct rte_flow_attr *attr,
3002                       const struct rte_flow_item items[],
3003                       const struct rte_flow_action actions[],
3004                       struct rte_flow_error *error)
3005 {
3006         struct rte_flow *flow = NULL;
3007         size_t size = 0;
3008         int ret;
3009
3010         ret = mlx5_flow_merge(dev, flow, size, attr, items, actions, error);
3011         if (ret < 0)
3012                 return NULL;
3013         size = ret;
3014         flow = rte_calloc(__func__, 1, size, 0);
3015         if (!flow) {
3016                 rte_flow_error_set(error, ENOMEM,
3017                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3018                                    NULL,
3019                                    "not enough memory to create flow");
3020                 return NULL;
3021         }
3022         ret = mlx5_flow_merge(dev, flow, size, attr, items, actions, error);
3023         if (ret < 0) {
3024                 rte_free(flow);
3025                 return NULL;
3026         }
3027         assert((size_t)ret == size);
3028         if (dev->data->dev_started) {
3029                 ret = mlx5_flow_apply(dev, flow, error);
3030                 if (ret < 0) {
3031                         ret = rte_errno; /* Save rte_errno before cleanup. */
3032                         if (flow) {
3033                                 mlx5_flow_remove(dev, flow);
3034                                 rte_free(flow);
3035                         }
3036                         rte_errno = ret; /* Restore rte_errno. */
3037                         return NULL;
3038                 }
3039         }
3040         TAILQ_INSERT_TAIL(list, flow, next);
3041         mlx5_flow_rxq_flags_set(dev, flow);
3042         return flow;
3043 }
3044
3045 /**
3046  * Create a flow.
3047  *
3048  * @see rte_flow_create()
3049  * @see rte_flow_ops
3050  */
3051 struct rte_flow *
3052 mlx5_flow_create(struct rte_eth_dev *dev,
3053                  const struct rte_flow_attr *attr,
3054                  const struct rte_flow_item items[],
3055                  const struct rte_flow_action actions[],
3056                  struct rte_flow_error *error)
3057 {
3058         return mlx5_flow_list_create
3059                 (dev, &((struct priv *)dev->data->dev_private)->flows,
3060                  attr, items, actions, error);
3061 }
3062
3063 /**
3064  * Destroy a flow in a list.
3065  *
3066  * @param dev
3067  *   Pointer to Ethernet device.
3068  * @param list
3069  *   Pointer to a TAILQ flow list.
3070  * @param[in] flow
3071  *   Flow to destroy.
3072  */
3073 static void
3074 mlx5_flow_list_destroy(struct rte_eth_dev *dev, struct mlx5_flows *list,
3075                        struct rte_flow *flow)
3076 {
3077         mlx5_flow_remove(dev, flow);
3078         TAILQ_REMOVE(list, flow, next);
3079         /*
3080          * Update RX queue flags only if port is started, otherwise it is
3081          * already clean.
3082          */
3083         if (dev->data->dev_started)
3084                 mlx5_flow_rxq_flags_trim(dev, flow);
3085         rte_free(flow);
3086 }
3087
3088 /**
3089  * Destroy all flows.
3090  *
3091  * @param dev
3092  *   Pointer to Ethernet device.
3093  * @param list
3094  *   Pointer to a TAILQ flow list.
3095  */
3096 void
3097 mlx5_flow_list_flush(struct rte_eth_dev *dev, struct mlx5_flows *list)
3098 {
3099         while (!TAILQ_EMPTY(list)) {
3100                 struct rte_flow *flow;
3101
3102                 flow = TAILQ_FIRST(list);
3103                 mlx5_flow_list_destroy(dev, list, flow);
3104         }
3105 }
3106
3107 /**
3108  * Remove all flows.
3109  *
3110  * @param dev
3111  *   Pointer to Ethernet device.
3112  * @param list
3113  *   Pointer to a TAILQ flow list.
3114  */
3115 void
3116 mlx5_flow_stop(struct rte_eth_dev *dev, struct mlx5_flows *list)
3117 {
3118         struct rte_flow *flow;
3119
3120         TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next)
3121                 mlx5_flow_remove(dev, flow);
3122         mlx5_flow_rxq_flags_clear(dev);
3123 }
3124
3125 /**
3126  * Add all flows.
3127  *
3128  * @param dev
3129  *   Pointer to Ethernet device.
3130  * @param list
3131  *   Pointer to a TAILQ flow list.
3132  *
3133  * @return
3134  *   0 on success, a negative errno value otherwise and rte_errno is set.
3135  */
3136 int
3137 mlx5_flow_start(struct rte_eth_dev *dev, struct mlx5_flows *list)
3138 {
3139         struct rte_flow *flow;
3140         struct rte_flow_error error;
3141         int ret = 0;
3142
3143         TAILQ_FOREACH(flow, list, next) {
3144                 ret = mlx5_flow_apply(dev, flow, &error);
3145                 if (ret < 0)
3146                         goto error;
3147                 mlx5_flow_rxq_flags_set(dev, flow);
3148         }
3149         return 0;
3150 error:
3151         ret = rte_errno; /* Save rte_errno before cleanup. */
3152         mlx5_flow_stop(dev, list);
3153         rte_errno = ret; /* Restore rte_errno. */
3154         return -rte_errno;
3155 }
3156
3157 /**
3158  * Verify the flow list is empty
3159  *
3160  * @param dev
3161  *  Pointer to Ethernet device.
3162  *
3163  * @return the number of flows not released.
3164  */
3165 int
3166 mlx5_flow_verify(struct rte_eth_dev *dev)
3167 {
3168         struct priv *priv = dev->data->dev_private;
3169         struct rte_flow *flow;
3170         int ret = 0;
3171
3172         TAILQ_FOREACH(flow, &priv->flows, next) {
3173                 DRV_LOG(DEBUG, "port %u flow %p still referenced",
3174                         dev->data->port_id, (void *)flow);
3175                 ++ret;
3176         }
3177         return ret;
3178 }
3179
3180 /**
3181  * Enable a control flow configured from the control plane.
3182  *
3183  * @param dev
3184  *   Pointer to Ethernet device.
3185  * @param eth_spec
3186  *   An Ethernet flow spec to apply.
3187  * @param eth_mask
3188  *   An Ethernet flow mask to apply.
3189  * @param vlan_spec
3190  *   A VLAN flow spec to apply.
3191  * @param vlan_mask
3192  *   A VLAN flow mask to apply.
3193  *
3194  * @return
3195  *   0 on success, a negative errno value otherwise and rte_errno is set.
3196  */
3197 int
3198 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
3199                     struct rte_flow_item_eth *eth_spec,
3200                     struct rte_flow_item_eth *eth_mask,
3201                     struct rte_flow_item_vlan *vlan_spec,
3202                     struct rte_flow_item_vlan *vlan_mask)
3203 {
3204         struct priv *priv = dev->data->dev_private;
3205         const struct rte_flow_attr attr = {
3206                 .ingress = 1,
3207                 .priority = MLX5_FLOW_PRIO_RSVD,
3208         };
3209         struct rte_flow_item items[] = {
3210                 {
3211                         .type = RTE_FLOW_ITEM_TYPE_ETH,
3212                         .spec = eth_spec,
3213                         .last = NULL,
3214                         .mask = eth_mask,
3215                 },
3216                 {
3217                         .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
3218                                 RTE_FLOW_ITEM_TYPE_END,
3219                         .spec = vlan_spec,
3220                         .last = NULL,
3221                         .mask = vlan_mask,
3222                 },
3223                 {
3224                         .type = RTE_FLOW_ITEM_TYPE_END,
3225                 },
3226         };
3227         uint16_t queue[priv->reta_idx_n];
3228         struct rte_flow_action_rss action_rss = {
3229                 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
3230                 .level = 0,
3231                 .types = priv->rss_conf.rss_hf,
3232                 .key_len = priv->rss_conf.rss_key_len,
3233                 .queue_num = priv->reta_idx_n,
3234                 .key = priv->rss_conf.rss_key,
3235                 .queue = queue,
3236         };
3237         struct rte_flow_action actions[] = {
3238                 {
3239                         .type = RTE_FLOW_ACTION_TYPE_RSS,
3240                         .conf = &action_rss,
3241                 },
3242                 {
3243                         .type = RTE_FLOW_ACTION_TYPE_END,
3244                 },
3245         };
3246         struct rte_flow *flow;
3247         struct rte_flow_error error;
3248         unsigned int i;
3249
3250         if (!priv->reta_idx_n) {
3251                 rte_errno = EINVAL;
3252                 return -rte_errno;
3253         }
3254         for (i = 0; i != priv->reta_idx_n; ++i)
3255                 queue[i] = (*priv->reta_idx)[i];
3256         flow = mlx5_flow_list_create(dev, &priv->ctrl_flows, &attr, items,
3257                                      actions, &error);
3258         if (!flow)
3259                 return -rte_errno;
3260         return 0;
3261 }
3262
3263 /**
3264  * Enable a flow control configured from the control plane.
3265  *
3266  * @param dev
3267  *   Pointer to Ethernet device.
3268  * @param eth_spec
3269  *   An Ethernet flow spec to apply.
3270  * @param eth_mask
3271  *   An Ethernet flow mask to apply.
3272  *
3273  * @return
3274  *   0 on success, a negative errno value otherwise and rte_errno is set.
3275  */
3276 int
3277 mlx5_ctrl_flow(struct rte_eth_dev *dev,
3278                struct rte_flow_item_eth *eth_spec,
3279                struct rte_flow_item_eth *eth_mask)
3280 {
3281         return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
3282 }
3283
3284 /**
3285  * Destroy a flow.
3286  *
3287  * @see rte_flow_destroy()
3288  * @see rte_flow_ops
3289  */
3290 int
3291 mlx5_flow_destroy(struct rte_eth_dev *dev,
3292                   struct rte_flow *flow,
3293                   struct rte_flow_error *error __rte_unused)
3294 {
3295         struct priv *priv = dev->data->dev_private;
3296
3297         mlx5_flow_list_destroy(dev, &priv->flows, flow);
3298         return 0;
3299 }
3300
3301 /**
3302  * Destroy all flows.
3303  *
3304  * @see rte_flow_flush()
3305  * @see rte_flow_ops
3306  */
3307 int
3308 mlx5_flow_flush(struct rte_eth_dev *dev,
3309                 struct rte_flow_error *error __rte_unused)
3310 {
3311         struct priv *priv = dev->data->dev_private;
3312
3313         mlx5_flow_list_flush(dev, &priv->flows);
3314         return 0;
3315 }
3316
3317 /**
3318  * Isolated mode.
3319  *
3320  * @see rte_flow_isolate()
3321  * @see rte_flow_ops
3322  */
3323 int
3324 mlx5_flow_isolate(struct rte_eth_dev *dev,
3325                   int enable,
3326                   struct rte_flow_error *error)
3327 {
3328         struct priv *priv = dev->data->dev_private;
3329
3330         if (dev->data->dev_started) {
3331                 rte_flow_error_set(error, EBUSY,
3332                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3333                                    NULL,
3334                                    "port must be stopped first");
3335                 return -rte_errno;
3336         }
3337         priv->isolated = !!enable;
3338         if (enable)
3339                 dev->dev_ops = &mlx5_dev_ops_isolate;
3340         else
3341                 dev->dev_ops = &mlx5_dev_ops;
3342         return 0;
3343 }
3344
3345 /**
3346  * Query flow counter.
3347  *
3348  * @param flow
3349  *   Pointer to the flow.
3350  *
3351  * @return
3352  *   0 on success, a negative errno value otherwise and rte_errno is set.
3353  */
3354 static int
3355 mlx5_flow_query_count(struct rte_flow *flow __rte_unused,
3356                       void *data __rte_unused,
3357                       struct rte_flow_error *error)
3358 {
3359 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
3360         if (flow->modifier & MLX5_FLOW_MOD_COUNT) {
3361                 struct rte_flow_query_count *qc = data;
3362                 uint64_t counters[2] = {0, 0};
3363                 struct ibv_query_counter_set_attr query_cs_attr = {
3364                         .cs = flow->counter->cs,
3365                         .query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
3366                 };
3367                 struct ibv_counter_set_data query_out = {
3368                         .out = counters,
3369                         .outlen = 2 * sizeof(uint64_t),
3370                 };
3371                 int err = mlx5_glue->query_counter_set(&query_cs_attr,
3372                                                        &query_out);
3373
3374                 if (err)
3375                         return rte_flow_error_set
3376                                 (error, err,
3377                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3378                                  NULL,
3379                                  "cannot read counter");
3380                 qc->hits_set = 1;
3381                 qc->bytes_set = 1;
3382                 qc->hits = counters[0] - flow->counter->hits;
3383                 qc->bytes = counters[1] - flow->counter->bytes;
3384                 if (qc->reset) {
3385                         flow->counter->hits = counters[0];
3386                         flow->counter->bytes = counters[1];
3387                 }
3388                 return 0;
3389         }
3390         return rte_flow_error_set(error, ENOTSUP,
3391                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3392                                   NULL,
3393                                   "flow does not have counter");
3394 #endif
3395         return rte_flow_error_set(error, ENOTSUP,
3396                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
3397                                   NULL,
3398                                   "counters are not available");
3399 }
3400
3401 /**
3402  * Query a flows.
3403  *
3404  * @see rte_flow_query()
3405  * @see rte_flow_ops
3406  */
3407 int
3408 mlx5_flow_query(struct rte_eth_dev *dev __rte_unused,
3409                 struct rte_flow *flow,
3410                 const struct rte_flow_action *actions,
3411                 void *data,
3412                 struct rte_flow_error *error)
3413 {
3414         int ret = 0;
3415
3416         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
3417                 switch (actions->type) {
3418                 case RTE_FLOW_ACTION_TYPE_VOID:
3419                         break;
3420                 case RTE_FLOW_ACTION_TYPE_COUNT:
3421                         ret = mlx5_flow_query_count(flow, data, error);
3422                         break;
3423                 default:
3424                         return rte_flow_error_set(error, ENOTSUP,
3425                                                   RTE_FLOW_ERROR_TYPE_ACTION,
3426                                                   actions,
3427                                                   "action not supported");
3428                 }
3429                 if (ret < 0)
3430                         return ret;
3431         }
3432         return 0;
3433 }
3434
3435 /**
3436  * Convert a flow director filter to a generic flow.
3437  *
3438  * @param dev
3439  *   Pointer to Ethernet device.
3440  * @param fdir_filter
3441  *   Flow director filter to add.
3442  * @param attributes
3443  *   Generic flow parameters structure.
3444  *
3445  * @return
3446  *   0 on success, a negative errno value otherwise and rte_errno is set.
3447  */
3448 static int
3449 mlx5_fdir_filter_convert(struct rte_eth_dev *dev,
3450                          const struct rte_eth_fdir_filter *fdir_filter,
3451                          struct mlx5_fdir *attributes)
3452 {
3453         struct priv *priv = dev->data->dev_private;
3454         const struct rte_eth_fdir_input *input = &fdir_filter->input;
3455         const struct rte_eth_fdir_masks *mask =
3456                 &dev->data->dev_conf.fdir_conf.mask;
3457
3458         /* Validate queue number. */
3459         if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
3460                 DRV_LOG(ERR, "port %u invalid queue number %d",
3461                         dev->data->port_id, fdir_filter->action.rx_queue);
3462                 rte_errno = EINVAL;
3463                 return -rte_errno;
3464         }
3465         attributes->attr.ingress = 1;
3466         attributes->items[0] = (struct rte_flow_item) {
3467                 .type = RTE_FLOW_ITEM_TYPE_ETH,
3468                 .spec = &attributes->l2,
3469                 .mask = &attributes->l2_mask,
3470         };
3471         switch (fdir_filter->action.behavior) {
3472         case RTE_ETH_FDIR_ACCEPT:
3473                 attributes->actions[0] = (struct rte_flow_action){
3474                         .type = RTE_FLOW_ACTION_TYPE_QUEUE,
3475                         .conf = &attributes->queue,
3476                 };
3477                 break;
3478         case RTE_ETH_FDIR_REJECT:
3479                 attributes->actions[0] = (struct rte_flow_action){
3480                         .type = RTE_FLOW_ACTION_TYPE_DROP,
3481                 };
3482                 break;
3483         default:
3484                 DRV_LOG(ERR, "port %u invalid behavior %d",
3485                         dev->data->port_id,
3486                         fdir_filter->action.behavior);
3487                 rte_errno = ENOTSUP;
3488                 return -rte_errno;
3489         }
3490         attributes->queue.index = fdir_filter->action.rx_queue;
3491         /* Handle L3. */
3492         switch (fdir_filter->input.flow_type) {
3493         case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
3494         case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
3495         case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
3496                 attributes->l3.ipv4.hdr = (struct ipv4_hdr){
3497                         .src_addr = input->flow.ip4_flow.src_ip,
3498                         .dst_addr = input->flow.ip4_flow.dst_ip,
3499                         .time_to_live = input->flow.ip4_flow.ttl,
3500                         .type_of_service = input->flow.ip4_flow.tos,
3501                         .next_proto_id = input->flow.ip4_flow.proto,
3502                 };
3503                 attributes->l3_mask.ipv4.hdr = (struct ipv4_hdr){
3504                         .src_addr = mask->ipv4_mask.src_ip,
3505                         .dst_addr = mask->ipv4_mask.dst_ip,
3506                         .time_to_live = mask->ipv4_mask.ttl,
3507                         .type_of_service = mask->ipv4_mask.tos,
3508                         .next_proto_id = mask->ipv4_mask.proto,
3509                 };
3510                 attributes->items[1] = (struct rte_flow_item){
3511                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
3512                         .spec = &attributes->l3,
3513                         .mask = &attributes->l3_mask,
3514                 };
3515                 break;
3516         case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
3517         case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
3518         case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
3519                 attributes->l3.ipv6.hdr = (struct ipv6_hdr){
3520                         .hop_limits = input->flow.ipv6_flow.hop_limits,
3521                         .proto = input->flow.ipv6_flow.proto,
3522                 };
3523
3524                 memcpy(attributes->l3.ipv6.hdr.src_addr,
3525                        input->flow.ipv6_flow.src_ip,
3526                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
3527                 memcpy(attributes->l3.ipv6.hdr.dst_addr,
3528                        input->flow.ipv6_flow.dst_ip,
3529                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
3530                 memcpy(attributes->l3_mask.ipv6.hdr.src_addr,
3531                        mask->ipv6_mask.src_ip,
3532                        RTE_DIM(attributes->l3_mask.ipv6.hdr.src_addr));
3533                 memcpy(attributes->l3_mask.ipv6.hdr.dst_addr,
3534                        mask->ipv6_mask.dst_ip,
3535                        RTE_DIM(attributes->l3_mask.ipv6.hdr.src_addr));
3536                 attributes->items[1] = (struct rte_flow_item){
3537                         .type = RTE_FLOW_ITEM_TYPE_IPV6,
3538                         .spec = &attributes->l3,
3539                         .mask = &attributes->l3_mask,
3540                 };
3541                 break;
3542         default:
3543                 DRV_LOG(ERR, "port %u invalid flow type%d",
3544                         dev->data->port_id, fdir_filter->input.flow_type);
3545                 rte_errno = ENOTSUP;
3546                 return -rte_errno;
3547         }
3548         /* Handle L4. */
3549         switch (fdir_filter->input.flow_type) {
3550         case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
3551                 attributes->l4.udp.hdr = (struct udp_hdr){
3552                         .src_port = input->flow.udp4_flow.src_port,
3553                         .dst_port = input->flow.udp4_flow.dst_port,
3554                 };
3555                 attributes->l4_mask.udp.hdr = (struct udp_hdr){
3556                         .src_port = mask->src_port_mask,
3557                         .dst_port = mask->dst_port_mask,
3558                 };
3559                 attributes->items[2] = (struct rte_flow_item){
3560                         .type = RTE_FLOW_ITEM_TYPE_UDP,
3561                         .spec = &attributes->l4,
3562                         .mask = &attributes->l4_mask,
3563                 };
3564                 break;
3565         case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
3566                 attributes->l4.tcp.hdr = (struct tcp_hdr){
3567                         .src_port = input->flow.tcp4_flow.src_port,
3568                         .dst_port = input->flow.tcp4_flow.dst_port,
3569                 };
3570                 attributes->l4_mask.tcp.hdr = (struct tcp_hdr){
3571                         .src_port = mask->src_port_mask,
3572                         .dst_port = mask->dst_port_mask,
3573                 };
3574                 attributes->items[2] = (struct rte_flow_item){
3575                         .type = RTE_FLOW_ITEM_TYPE_TCP,
3576                         .spec = &attributes->l4,
3577                         .mask = &attributes->l4_mask,
3578                 };
3579                 break;
3580         case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
3581                 attributes->l4.udp.hdr = (struct udp_hdr){
3582                         .src_port = input->flow.udp6_flow.src_port,
3583                         .dst_port = input->flow.udp6_flow.dst_port,
3584                 };
3585                 attributes->l4_mask.udp.hdr = (struct udp_hdr){
3586                         .src_port = mask->src_port_mask,
3587                         .dst_port = mask->dst_port_mask,
3588                 };
3589                 attributes->items[2] = (struct rte_flow_item){
3590                         .type = RTE_FLOW_ITEM_TYPE_UDP,
3591                         .spec = &attributes->l4,
3592                         .mask = &attributes->l4_mask,
3593                 };
3594                 break;
3595         case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
3596                 attributes->l4.tcp.hdr = (struct tcp_hdr){
3597                         .src_port = input->flow.tcp6_flow.src_port,
3598                         .dst_port = input->flow.tcp6_flow.dst_port,
3599                 };
3600                 attributes->l4_mask.tcp.hdr = (struct tcp_hdr){
3601                         .src_port = mask->src_port_mask,
3602                         .dst_port = mask->dst_port_mask,
3603                 };
3604                 attributes->items[2] = (struct rte_flow_item){
3605                         .type = RTE_FLOW_ITEM_TYPE_TCP,
3606                         .spec = &attributes->l4,
3607                         .mask = &attributes->l4_mask,
3608                 };
3609                 break;
3610         case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
3611         case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
3612                 break;
3613         default:
3614                 DRV_LOG(ERR, "port %u invalid flow type%d",
3615                         dev->data->port_id, fdir_filter->input.flow_type);
3616                 rte_errno = ENOTSUP;
3617                 return -rte_errno;
3618         }
3619         return 0;
3620 }
3621
3622 /**
3623  * Add new flow director filter and store it in list.
3624  *
3625  * @param dev
3626  *   Pointer to Ethernet device.
3627  * @param fdir_filter
3628  *   Flow director filter to add.
3629  *
3630  * @return
3631  *   0 on success, a negative errno value otherwise and rte_errno is set.
3632  */
3633 static int
3634 mlx5_fdir_filter_add(struct rte_eth_dev *dev,
3635                      const struct rte_eth_fdir_filter *fdir_filter)
3636 {
3637         struct priv *priv = dev->data->dev_private;
3638         struct mlx5_fdir attributes = {
3639                 .attr.group = 0,
3640                 .l2_mask = {
3641                         .dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
3642                         .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
3643                         .type = 0,
3644                 },
3645         };
3646         struct rte_flow_error error;
3647         struct rte_flow *flow;
3648         int ret;
3649
3650         ret = mlx5_fdir_filter_convert(dev, fdir_filter, &attributes);
3651         if (ret)
3652                 return ret;
3653         flow = mlx5_flow_list_create(dev, &priv->flows, &attributes.attr,
3654                                      attributes.items, attributes.actions,
3655                                      &error);
3656         if (flow) {
3657                 DRV_LOG(DEBUG, "port %u FDIR created %p", dev->data->port_id,
3658                         (void *)flow);
3659                 return 0;
3660         }
3661         return -rte_errno;
3662 }
3663
3664 /**
3665  * Delete specific filter.
3666  *
3667  * @param dev
3668  *   Pointer to Ethernet device.
3669  * @param fdir_filter
3670  *   Filter to be deleted.
3671  *
3672  * @return
3673  *   0 on success, a negative errno value otherwise and rte_errno is set.
3674  */
3675 static int
3676 mlx5_fdir_filter_delete(struct rte_eth_dev *dev __rte_unused,
3677                         const struct rte_eth_fdir_filter *fdir_filter
3678                         __rte_unused)
3679 {
3680         rte_errno = ENOTSUP;
3681         return -rte_errno;
3682 }
3683
3684 /**
3685  * Update queue for specific filter.
3686  *
3687  * @param dev
3688  *   Pointer to Ethernet device.
3689  * @param fdir_filter
3690  *   Filter to be updated.
3691  *
3692  * @return
3693  *   0 on success, a negative errno value otherwise and rte_errno is set.
3694  */
3695 static int
3696 mlx5_fdir_filter_update(struct rte_eth_dev *dev,
3697                         const struct rte_eth_fdir_filter *fdir_filter)
3698 {
3699         int ret;
3700
3701         ret = mlx5_fdir_filter_delete(dev, fdir_filter);
3702         if (ret)
3703                 return ret;
3704         return mlx5_fdir_filter_add(dev, fdir_filter);
3705 }
3706
3707 /**
3708  * Flush all filters.
3709  *
3710  * @param dev
3711  *   Pointer to Ethernet device.
3712  */
3713 static void
3714 mlx5_fdir_filter_flush(struct rte_eth_dev *dev)
3715 {
3716         struct priv *priv = dev->data->dev_private;
3717
3718         mlx5_flow_list_flush(dev, &priv->flows);
3719 }
3720
3721 /**
3722  * Get flow director information.
3723  *
3724  * @param dev
3725  *   Pointer to Ethernet device.
3726  * @param[out] fdir_info
3727  *   Resulting flow director information.
3728  */
3729 static void
3730 mlx5_fdir_info_get(struct rte_eth_dev *dev, struct rte_eth_fdir_info *fdir_info)
3731 {
3732         struct rte_eth_fdir_masks *mask =
3733                 &dev->data->dev_conf.fdir_conf.mask;
3734
3735         fdir_info->mode = dev->data->dev_conf.fdir_conf.mode;
3736         fdir_info->guarant_spc = 0;
3737         rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
3738         fdir_info->max_flexpayload = 0;
3739         fdir_info->flow_types_mask[0] = 0;
3740         fdir_info->flex_payload_unit = 0;
3741         fdir_info->max_flex_payload_segment_num = 0;
3742         fdir_info->flex_payload_limit = 0;
3743         memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
3744 }
3745
3746 /**
3747  * Deal with flow director operations.
3748  *
3749  * @param dev
3750  *   Pointer to Ethernet device.
3751  * @param filter_op
3752  *   Operation to perform.
3753  * @param arg
3754  *   Pointer to operation-specific structure.
3755  *
3756  * @return
3757  *   0 on success, a negative errno value otherwise and rte_errno is set.
3758  */
3759 static int
3760 mlx5_fdir_ctrl_func(struct rte_eth_dev *dev, enum rte_filter_op filter_op,
3761                     void *arg)
3762 {
3763         enum rte_fdir_mode fdir_mode =
3764                 dev->data->dev_conf.fdir_conf.mode;
3765
3766         if (filter_op == RTE_ETH_FILTER_NOP)
3767                 return 0;
3768         if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
3769             fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
3770                 DRV_LOG(ERR, "port %u flow director mode %d not supported",
3771                         dev->data->port_id, fdir_mode);
3772                 rte_errno = EINVAL;
3773                 return -rte_errno;
3774         }
3775         switch (filter_op) {
3776         case RTE_ETH_FILTER_ADD:
3777                 return mlx5_fdir_filter_add(dev, arg);
3778         case RTE_ETH_FILTER_UPDATE:
3779                 return mlx5_fdir_filter_update(dev, arg);
3780         case RTE_ETH_FILTER_DELETE:
3781                 return mlx5_fdir_filter_delete(dev, arg);
3782         case RTE_ETH_FILTER_FLUSH:
3783                 mlx5_fdir_filter_flush(dev);
3784                 break;
3785         case RTE_ETH_FILTER_INFO:
3786                 mlx5_fdir_info_get(dev, arg);
3787                 break;
3788         default:
3789                 DRV_LOG(DEBUG, "port %u unknown operation %u",
3790                         dev->data->port_id, filter_op);
3791                 rte_errno = EINVAL;
3792                 return -rte_errno;
3793         }
3794         return 0;
3795 }
3796
3797 /**
3798  * Manage filter operations.
3799  *
3800  * @param dev
3801  *   Pointer to Ethernet device structure.
3802  * @param filter_type
3803  *   Filter type.
3804  * @param filter_op
3805  *   Operation to perform.
3806  * @param arg
3807  *   Pointer to operation-specific structure.
3808  *
3809  * @return
3810  *   0 on success, a negative errno value otherwise and rte_errno is set.
3811  */
3812 int
3813 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
3814                      enum rte_filter_type filter_type,
3815                      enum rte_filter_op filter_op,
3816                      void *arg)
3817 {
3818         switch (filter_type) {
3819         case RTE_ETH_FILTER_GENERIC:
3820                 if (filter_op != RTE_ETH_FILTER_GET) {
3821                         rte_errno = EINVAL;
3822                         return -rte_errno;
3823                 }
3824                 *(const void **)arg = &mlx5_flow_ops;
3825                 return 0;
3826         case RTE_ETH_FILTER_FDIR:
3827                 return mlx5_fdir_ctrl_func(dev, filter_op, arg);
3828         default:
3829                 DRV_LOG(ERR, "port %u filter type (%d) not supported",
3830                         dev->data->port_id, filter_type);
3831                 rte_errno = ENOTSUP;
3832                 return -rte_errno;
3833         }
3834         return 0;
3835 }